⚡ Local Inference Recipes

English | 简体中文 | 日本語

DeepSeek: local inference speed (tok/s), sourced

As of 2026-11-14, this dataset has 151 recipes for DeepSeek. 96 were measured by the poster. Fastest measured single-device, single-stream text result: 86.1 tok/s with DeepSeek-V4.1-Flash-1017 on Strata (1× RTX PRO 6000 Blackwell Workstation 96GB + Ryzen 7 9800X3D + 61GB RAM (Linux); Measured, 2026-10-26; source). 138 entries have a date, from 2026-07-31 to 2026-11-14.

Filter these in the explorer →

ModelEngineHardwaretok/sMetricSource levelDateSource
DeepSeek-V4.1-Flash-1017spark-ds41f4× DGX Spark (GB10)136Code decode 136; prefill of a 256K-class prompt 6,221; prefill of a 1M prompt 4,892; the tester says decode is about 8–13% faster and prefill about 2.6–2.7 times faster than TensorFold under the same conditions; on 2026-11-03 _m_naoki_m_ used the same spark-ds41f recipe (vLLM + TileLang, TP4, 200GbE RoCE switched mesh): prefill at 32K about 7,127, aggregate (separate) C16 about 235.5 (x.com/_m_naoki_m_/status/2117429844824977818); Another tester (masa_ai_labo) with spark-ds41f-1030: 4× Spark TP4, 256K prose, 10-run average decode 74.73, cold prefill 6,441 (x.com/masa_ai_labo/status/2117875287384097030); on 2026-11-07 the engine author christopherowen with spark-ds41f-1030 (4× Spark, no switch) reported single-stream 78.6 (95% CI 75.3–80.2); aggregate (separate) C8 total 198.7 / C16 total 258.1 (x.com/christopherowen/status/2119038167047094467). This does not replace this row's code decode 136, and does not replace comm-084 at 91.3 or x-190 at 92. The spark-ds41f-1107 author reports 4 Sparks, native 1017 weights and 1M limit: single-stream 82.5, C16 total 273.2, prefill 8661 at 131K and 6015 at 1M (x.com/christopherowen/status/2119210831812370771). kevintheo's separate 4 Spark coding corpus reaches C16 total 361 (x.com/kevintheo/status/2119363056333476146); tasks differ and rates are not interchangeable. Keep this row's 136 and do not alter comm-084 or x-190. On 2026-11-10 kevintheo used christopherowen spark-ds41f-1107 on 4 Sparks in a switchless ring: about 130 on 1 coding prompt, and a 16-prompt coding total of 361; 130 is within 5% of this row's code 136, so this is a note only. The same day JimHallers on 1107 (built from main@51be7406) reports 32K prefill 2,700 to 8,141 and a 16-stream total 141 to 257; the author says decode is level with the old stack. Keep 136.Measured2026-11-01link
DeepSeek-V4.1-Flash-1017TensorFold3× DGX Spark (TP3)126.1LakyFx recipe 1.2.5 (K2a, TensorFold TP3; 1.2.5 is the recipe version, not a TensorFold version), SparkDash: C1 code 126.1 / prose 87.7, tok_s is code; aggregate (separate) C12 code 246.2; recipe update on 2026-10-31: TP3 prefill about 2,200 → about 3,500, and a TP4 recipe is added (x.com/laky_fx/status/2116433866340569124)Measured2026-10-23link
DeepSeek-V4.1-FlashTensorFold4× DGX Spark (GB10, switchless ring)111code, single stream, median of 3: stock SGLang 71.6, Mia SGLang branch 103, TensorFold 111; aggregate (separate) 4 streams 71.5/116/140. Author handle corrected to kevintheo (X user ID 952585832115920897; the earlier kevintheoo was a typo and that handle does not exist).Measured2026-10-26link
DeepSeek-V4.1-Flash-1017vLLM2× DGX Spark (GPU locked at 2300 MHz) + 1 EPYC host with 2× RTX 3090 (NVLink), 3 machines in total110sparkDash protocol, median of 3, single stream: JSON 138, code 110, prose (temperature 0, thinking off) 66, prose (thinking on) 56; aggregate (separate) 4 streams: 322 / 192 / 130 / 105. Prefill 4K 2,827, 128K 2,650, 256K 2,618. Decode after a 128K prompt 62Measured2026-10-31link
DeepSeek-V4.1-Flash-1017TensorFold4× DGX Spark (GB10, RoCE)97.2New lane beyond prompt set v1 (recipe v0.3.5, engine 0a0b69a + 2 patches, both RoCE links): 1 user 97.2 (code single stream 143.7); aggregate (separate) 6 users 195.5 / 16 users 230.7; a 93K prompt is read in about 37 s; KV pool 29,360,128 (was 2,097,152); 600K needle passed. Previous lane (LibertAI EXL3 3.5 bpw): C1 89.8 / 93.6 (with the second RoCE rail), C6 aggregate 202.9 / 213.8, cold prefill 93K 1,929 / 2,215Measured2026-10-27link
DeepSeek-V4.1-FlashTensorFold2× DGX Spark95TensorFold: 1 stream 95, 3 streams 147, prefill 1,400; vLLM comparison: 51 / 86 / 870Measured2026-10-03link
DeepSeek-V4.1-Flash-1017TensorFold4× DGX Spark (GB10) + MikroTik CRS804-4DDQ switch (1 unit)95single stream at 32k depth: code about 95–115 (tok_s is the low end, 95), prose about 67; prefill about 3k (32k prompt in 11 s); 1M contextMeasured2026-10-26link
DeepSeek-V4.1-Flash-1017vLLM3× DGX Spark + 1× Acer GN100 (4 machines in a ring)921 chat 92–104; the author's own EXL3 4.0bpw (275GB) on the same machines 83–88; time to first token on a 128K prompt 18.8 s vs 64.4 s; at 300K 48.0 s vs 158.3 s; aggregate (separate) 16 chats 348 vs 202.1Measured2026-10-30link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (direct QSFP link)88.7one user, median of 5, temp 0.7: json 102.8 / code 88.7 / explain 72.8 / story 57.3, tok_s is code; 8K prompt read in 3.0 sMeasured2026-10-24link
DeepSeek-V4.1-Flash-1017Strata1× RTX PRO 6000 Blackwell Workstation 96GB + Ryzen 7 9800X3D + 61GB RAM (Linux)86.1short decode 86.1 (58.5–130.9); 32K prefill 2,185.6; Linux. For Windows data on the same card, see the other entryMeasured2026-10-26link
DeepSeek-V4.1-Flash-1017NVIDIA Flare4× DGX Spark (QSFP ring, no switch)85single stream: counting 92–98 / code 85–89; 62.5 over a 114K-token Max-effort output; author says faster than their vLLM recipe at 1–3 streamsMeasured2026-10-23link
DeepSeek-V4.1-Flash-1017pico-dgx-ds44× DGX Spark (GB10) TP4, 200G RoCE switch82pico-dgx-ds4 at 7a3b982 (blaesus/dsgx10 fork, DSGB10). Recipe-author README headline table, temperature 0, 9-run median: single-user chat decode 82; code 144 is separate. sparkDash 1.8.8 method: prose 101 / code 166. Multi-user totals of 300 (8 distinct prompts), 380 (16 distinct) and 438 (16 copied prompts) are not single-stream. Headline is the exact/lossless build; a lossy relaxed-accept trial was removed. After images: chat/code 81.9/144.4. --max-ctx 1048576. The table does not say per row whether the official checkpoint or the UNCENSORED-FP8 variant was used. 2026-11-11 same-recipe live build (stack4 + stack5): 27-run median on the same 3-prompt chat bench is 84.6, code 144.0, 3.2% above the existing 82 and within ±5%; appended only, headline unchanged, no new row. 90.3 is the mean on a different 320-prompt held-out set (exact 86.2 → stack4 89.5 → stack5 90.3), a different method. 16-user distinct-prompt totals 368–376 are not single-stream. Date and source URL are unchanged. https://github.com/Picom-code/pico-dgx-ds4Measured2026-11-09link
DeepSeek-V4-Flash (0731 GA)Unspecified2× DGX Spark82single stream 82 tok/s (peak about 95); 3 sessions 135 tok/sMeasured2026-07-31link
DeepSeek-V4.1-Flash-1017Unspecified2× Radeon AI PRO R9700 32GB + 128GB RAM82About 82–91 with a warm cache. The tester says this is real use, not a quick benchmark. The post does not name the engineMeasured2026-11-03link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (GB10, CX7)80.7sparkDash (CONTEXT=524288, drafts on): single-stream prose 74.9 / code 109.9; 4-stream aggregate prose 130.0 / code 155.5; cold prefill 8K–256K about 2,000–2,183; weight card is experimental with no PPL/KL vs unquantized; same-class Elevatormusic (comm-038) prose 56.8; yume_arasaki corrected their re-test on 10-23: prose 70+ (not 58), code about 107; wei_wang repeated the Mia two-Spark recipe on 2026-11-01 (EXL3 2.9 bpw, TP=2): single-stream code 118 (x.com/wei_wang/status/2116925488698449964). 2026-11-09 Mia README current sparkDash table (CONTEXT=524288, drafts on, temperature 0, thinking off): single-stream prose 80.7, headline updated; code 118.1 is separate. Two-stream prose total 110.6 / four-stream total 154.0 are not single-stream. Recipe-author first-hand table. https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-SparksMeasured2026-11-09link
DeepSeek-V4.1-Flash-1017TensorFoldMac Studio M5 Ultra 256GB80.5code 80.5, prose 70.5; first token after a 32K prompt 18.6 s; aggregate (separate) 8 requests 88.3Measured2026-10-26link
DeepSeek-V4.1-Flash-1017TensorFold4× DGX Spark (GB10, RoCE)79.1recipe v0.2.0 set b median: single-stream code 135.9 / prose 79.1 / structured 186.6; 4-stream aggregate 141.7, 4-stream sustained 90s aggregate 183.1; cold prefill 8K 2,793, 128K 2,679; 4.02bpw is about 15–17% slower than v0.1.1 2.9bpw (pre-1017 Mia weights); KL not measured; X post quotes code 136 / 4-stream 183; 10-23 check: README v0.3.3 single-stream code 135.6 / prose 78.6, 16-stream sustained aggregate 255.6; Gael_DDOS re-test prose 78.6Measured2026-10-19link
DeepSeek-V4-Flash (0731)affinityThreadripper PRO 9955WX + 128GB + 4× Radeon AI PRO R970079greedy decode 79 tok/s (63 at temp 1.0; ~90 on chat with thinking)Measured2026-10-16link
DeepSeek-V4.1-Flash-1017spark-ds41f4× DGX Spark class GB10 (2 ASUS Ascent GX10, Gigabyte AI TOP ATOM, Lenovo ThinkStation PGX), TP4, CX7 cable ring (no switch)76.3Owen's spark-ds41f standard release 1107 (release commit cc20ed18), vLLM fork tree ed1bdb7a + TileLang, TP4 cable ring, DSpark 5 drafts, 1,048,576 context, 16 slots. Deviations from the recipe: 4 KiB page kernel, no memory saver, 2,200 MHz GPU clock cap, image built locally. Thinking off, temperature 1.0, 512 tokens, decode is the median of 2 runs each. Headline is prose decode 76.3; code 76.4. Context-step decode (prose) at 16K / 32K / 64K / 128K: 76.7 / 81.0 / 76.7 / 74.1. Prefill 32K 8,355, 128K 7,964. The 8-request aggregate of 201.4 is not a single-stream value. A repeat the same night gave decode 81.5. Same test on candidate-54: prose 79.9 / code 89.1; the author says the difference is inside run-to-run spread. The result page has speed2.json, logs and commands https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-owen-ds41f-1107-tp4-ring/ , so this is measured. The same author's TensorFold 4x Spark row is x-199.Measured2026-11-13link
DeepSeek-V4.1-Flash-1017-AbliteratedoMLXMac Studio M5 Ultra (64-core GPU) 256GB75Decode 75, prefill 1,600 (300 with default settings; settings were tuned); Qwen3.8-27B on the same Mac 60 (not tuned)Measured2026-10-31link
DeepSeek-V4.1-Flash-1017TensorFold3× DGX Spark (TP3)74.7TensorFold TP3: aggregate 108.6 over 4 sessions (SGLang TP3 68.3), up to 242.9 with 4 streams; no single-stream figure; on 2026-11-01 the tester added a single-stream figure of 74.7 (German, thinking on; x.com/HandwerkAiLab/status/2116961530550095915)Measured2026-10-24link
DeepSeek-V4.1-Flash-1017TensorFold3× DGX Spark (GB10, 3 direct CX7 cables, no switch; clock capped at 2000 MHz)73.8single-stream set b: code 122.6 / prose 73.8 / structured 147.6; set a: 111.9 / 67.5 / 129.6; 2/4/8-stream aggregate 93–102 / 121–136 / 165–181; prefill 8K/32K/128K 2,222/2,257/2,074; same author 2.9bpw post: TP3 vs TP2 code 117 vs 88, prose 82 vs 61Measured2026-10-23link
DeepSeek-V4.1-Flash-1017TensorFold4× DGX Spark (TP4)71Mia version of Nov 1 (commit fcef6cf, KV pool 2,097,152, 16 request slots): code 71.0, prose 68.4, HTML answer 99.8; first token at 8K 2.2 s; aggregate (separate) 1 / 8 at once 62.1 / 144.6. vLLM on the same boxes (ondigo-winder recipe): code 33.6, prose 30.2, 8 at once 114.8. Old value: code 69.5 on 10-30Measured2026-11-02link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (one is an ASUS GX10; direct QSFP link, TP2)70.5bertholomus recipe v0.7.2 @30a5bf9: decode code 70.5 / prose 70.6; 32K prefill 1,830, TTFT 17.9 s; aggregate (separate) 4 requests 107.4, 8 requests 106.3 (the recipe has 4 slots); for comparison 4× Spark 4.0bpw: 82.8 / 80.8Measured2026-10-25link
DeepSeek-V4.1-FlashProject Maya4× RTX 3090 + EPYC (Zen 2)70decode 57 (stock 33); 32K prefill 1,626→2,642; about 30% of experts in RAM; on 2026-11-01, after all 8 memory channels were filled (512 GB, 44→72 GB/s), a short code replay with Maya-S went 67.7→81.5 (1 run each, x.com/superalesha/status/2116937129569738863); New post on 2026-11-03: 4× 3090 (PCIe 3.0, no NVLink), decode about 70 on stock Maya v1.1.51 and on the author build; author build prefill 16K 1,029 → 1,262, 116K 1,753 → 2,651, 272K 1,406 → 2,888; previous main value 57 (x.com/superalesha/status/2114402413742362907)Measured2026-11-03link
DeepSeek-V4.1-Flash-1017Project Maya2× RX 7900 XTX 24GB TP2 (PCIe 4.0 x4 each, −100mV, 350W cap) + Ryzen AI Max+ 395 128GB (8060S as helper)69.4Project Maya v1.1.56 (22fc101f) + open PRs #867/#904/#945/#967/#968 (tag maya-ds4-exl3-1m-20261112 = 3c4e104). A public reachable commit counts as a version. --bench decode was 69.4 on both loads; that is the headline. --max-context 1048576, DSpark on. Long-prompt 64-token retrieval-copy up to 85.8 is a structured task, not the headline. #967 helper sums are not bit-identical to the cards (author note). Older values on the same machine: 2026-11-07, v1.1.56 + PRs #867/#904 (recipe f5aa4cb, quant rev 9183eaa), --bench prose decode 59.6; prefill 1,656 at 160K and 651.5 at 8K; code at 130K 50.3; the 64-token retrieval-copy peak 79.2 / 67.4 at 1.04M is a structured task, not the headline (https://github.com/mw00/project-maya/issues/905 ). In the same issue #969, v1.1.56 alone (older stock recipe from #871) decoded 49.8. --bench table and recipe are public.Measured2026-11-12link
DeepSeek-V4.1-Flash-1017TensorFoldMac Studio M5 Ultra 80C 256GB67.5recipe-harness single stream (drafts on, thinking off, greedy): prose 67.5 / code 87.5 / structured 77.6 / list 94.4; 4-stream aggregate prose 93.5 / code 101.4; cold prefill ~2k–131k (thinking off) about 1,093–1,256; 1M needle passed; repo name says mlx, body is TensorFold Metal. 2026-11-13: tanchunsiong reports decode 126-147 with the same BobClawblaw recipe (TensorFold native Metal, Mac Studio M5 Ultra 256GB). The TensorFold version, prompt type and drafter settings are not stated, so the basis is unclear; note only, the headline does not change. https://x.com/tanchunsiong/status/2121072640773574980Measured2026-10-19link
DeepSeek-V4.1-Flash-1017TensorFold3× DGX Spark (GB10 128GB) TP366TensorFold TP3 integration recipe (README bench table at 461d8cd; later 02b133f is a scheduling fix and is after the bench). Recipe-author table: single-request prose decode 66; code 101 and structured 147 are separate. Short synthetic prompts, thinking off, 384-token cap. Multi-client C4/C8/C12 totals 154/201/222 are not single-stream; C12 per-request median 23.4. The 6.4M figure is a shared KV pool, not a per-request limit; context is left null. Max output 65,536.Measured2026-11-10link
DeepSeek-V4.1-Flash-1017vLLM2× DGX Spark (TP2)64.9sudoingX vLLM recipe v2.0 @f8f20b3, separate speed run with thinking off: code 64.9 / prose 62.2; aggregate (separate) 8 requests 66.3Measured2026-10-23link
DeepSeek-V4-FlashvLLM2× DGX Spark (QSFP direct link)63.7single stream 63.7 tok/s (bandwidth-limit estimate 74.7, reaches 85%); V4.1-Flash on same machine failed due to RAMMeasured2026-09-28link
DeepSeek-V4.1-FlashSGLang4× DGX Spark (100 GbE switch)62.5single-stream prose ~62.5 tok/s (tuned; base ~32.6; vs Moorhead ~58; at 16 slots prose ~228 / code ~293; media 494 not reached)Measured2026-10-16link
DeepSeek-V4.1-Flash-1017SGLang4× DGX Spark (GB10, 200G RoCE)62Single-stream decode: code 106 / prose 62 tok/s; 8-way aggregate 190–290; prefill ~5,300–5,500 (8K–128K), ~4,850 at 240K; TTFT ~0.22 s short prompt, ~24 s at 128KMeasured2026-10-18link
DeepSeek-V4.1-Flash-1017Project Maya2× RX 7900 XTX 24GB TP2 (PCIe 4.0 x4 each, −100mV, 350W cap) + Ryzen AI Max+ 395 128GB (8060S as helper)60.1Same machine and --bench table as comm-366, v1.1.56, no extra PRs. GSQ-RCO IQ2_XS decode 60.1, KL 0.227. Configured 1,048,576. Different quant, so this is a new row.Measured2026-11-12link
DeepSeek-V4.1-Flash-1017vLLM2× DGX Spark59.2v1.2 set b: prose 59.2 / code 100.7 / structured 141.3 (same session launch build 58.0 / 103.0 / 142.6); KLD vs native 1017 0.0243, top-1 95.47% (same panel LibertAIDAI 2.9 is KLD 0.0304); v1.2 prefill not re-measured; v1.1 8K/32K 2,090/2,092 in x-059Measured2026-10-20link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (GB10, both fabric links)59.1Tuned profile, thinking off, temp 0, 400-token output cap, median of two isolated passes: C1 (single stream) prose 59.1 / code 103.4 / structured counting 91.0; aggregate C2/C4/C8: code 149.7 / 193.3 / 267.0, prose 85.0 / 110.1 / 139.1; cold prefill 72K 929; earlier comparison (64-token output) C1 decode TensorFold 58.1 vs vLLM 36.9Measured2026-10-18link
DeepSeek-V4.1-Flash (sudoingX 2.7-bit mixed-bit-width experts)vLLM2× DGX Spark57.9v1.1: code 103.0 / prose 57.9 / structured 141.0 tok/s (DSpark on); prefill 8K 2,090, 32K 2,092 tok/s (v1 1,763 / 1,773); 32K first token 15.7 sMeasured2026-10-13link
DeepSeek-V4.1-Flash (sudoingX 2.7-bit mixed-bit-width experts)vLLM2× DGX Spark57.2reproduces sudoingX recipe, single stream thinking off, 3 runs <1% difference: structured 135.5 / JSON 115.2 / math 114.0 / code 91.3 / prose 57.2; draft off about 44; 8K→240K prose 57.6–61.5; TTFT 8K 4.3 s / 35K ~18 s / 240K 91.8 s; Hermes Agent turns 82–94Measured2026-10-15link
DeepSeek-V4.1-Flash-1017 BLACKFOLD (refusal-removed)TensorFold2× DGX Spark (GB10, two RoCE rails)57.1125K-context A/B (4 reps, temp 0): single-stream code 92.8 / prose 57.1 (Python engine 89.6 / 55.2); 2 streams code 69.8 each; 125K cold prefill 62.0 s; outputs token-identical to reference; author says prose is level with lucatacoMeasured2026-10-20link
DeepSeek-V4.1-Flash-1017vLLM2× DGX Spark (GB10; GPU clock locked 2000 MHz)57prose about 57 tok/s (1 stream) / aggregate about 91 (3 streams); Elevatormusic same scripts: prose 50.6, code 82.4–88.9 (vs TF 56.8/92.2, TF still faster); KLD 0.0201 / top-1 95.96%; cold 314K prefill 2,066; this run locked at 2000 MHz; Elevatormusic logs ~2.4 GHz (their post does not pin a clock)Measured2026-10-19link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (GB10, direct CX7 link)56.8Single-stream decode (median of 6 engine builds, thinking off): prose 56.8 / code 92.2; thinking on 54.2 / 71.6; vLLM (MiaAI kit) on the same pair and weights 27.0 / 39.9; aggregate (released engine) prose 73.5 / 82.3 / 94.4 at 2/3/4 streams, code 140.3 at 4; cold prefill 32K 2,165 (median of 5, 2,065–2,201); needle passed at 1,035,020 tokens (TTFT 13.6 min)Measured2026-10-18link
DeepSeek-V4.1-Flash-1017TensorFold2× DGX Spark (GB10, CX7)55.48 slots, 400K pool, vision tower on: 1 stream prose 55.4 / code 103.5 / structured 62.7 / list 70.9; aggregate 2 streams 81.5 / 126.2, 4 streams 108.8 / 163.8, 8 streams 134.8 / 202.7 (prose / code); first text 0.19–0.26 s; without vision tower code 98.1 / 158.5 at 1 / 4 streamsMeasured2026-10-18link
DeepSeek-V4.1-Flash-1017 (abliterated)oMLX1× Mac Studio M5 Ultra 256GB55.3oMLX 0.7.4.dev1 full bench table: pp1K decode 55.3 / prefill 640.8; pp4K decode 62.1 / prefill 1,306. Same post: two M5 Ultras over TB5 RDMA prefills worse than one Mac; enabling CEP/INT8 also slows decode. tok_s uses pp1K 55.3Measured2026-11-05link
DeepSeek-V4.1-Flash-1017TensorFold2× ASUS Ascent GX10 (GB10)55.2single-stream code 92.9 / JSON 104.3 / prose 55.2; first token 0.23 s; GPU draw about 78W for both boxes during code runMeasured2026-10-20link
DeepSeek-V4.1-Flash-1017ds4 (DwarfStar)Mac Studio M5 Ultra 256GB54.5kernelpool ds4 fork @72ccf36, batched mode with 5 slots: single stream code 54.5 / prose 54.5; aggregate (separate) 4 requests 56.2, 8 requests 50.5; 32K prefill 1,475Measured2026-10-24link
DeepSeek-V4.1-Flash-1017ExLlama2× DGX Spark50.3By draft depth: no draft 31.3; prose best 50.3 (depth 2); code best 87.3 (depth 6); structured-task best 131.6 (depth 5, structured output only, not a general speed). tok_s is proseMeasured2026-10-28link
DeepSeek-V4.1-FlashvLLM2× DGX Spark48.535K agent prompt 48.5 tok/s; ≈30–32.6 at 17K–559K depthMeasuredlink
DeepSeek-V4.1-Flash-1017Project Maya1× RTX PRO 5000 72GB (eGPU, current PCIe x4) + Ryzen AI Max+ 395 128GB (CachyOS)48.3Project Maya v1.1.54, DSpark on, --max-context 262144. Headline is the author's attached official --bench: greedy, thinking off, mean 48.3 over three answers of up to 256 tokens (code 47.1 / explain 43.5 / HTML 54.2); prefill 150 at 2K and 506 at 8K. The issue form and screenshot say decode 70 and prefill 862, which do not match --bench; live answers in the log around 63–73 are not the headline. The form's “prompt length 500k” conflicts with the 256K limit and is not used. maya-bench.txt and logs are in the comment attachments.Measured2026-11-12link
DeepSeek-V4.1-Flash-1017Project Maya1× RX 7900 XTX 24GB (Radeon 8060S helper) + Ryzen AI Max+ 395 + 128GB RAM45.6Maya 1.1.56 (22fc101f) + PR 904, build f5aa4cb. One XTX on PCIe 4.0 x4; Radeon 8060S helper on the Ryzen AI Max+ 395 host has 128GB shared RAM and a 113GB helper pool. --bench prose decode 45.6; stock 43.3; context limit 1048576, 2K/8K benchmark prompts. Structured retrieval peak 55.7 and 51.1 at 1.04M are not the headline. About 29GB of experts remain file-backed; short prompts pay about 16s of fixed I/O. Different GPU count from comm-366.Measured2026-11-07link
DeepSeek-V4.1-Flash-1017 (abliterated)Project MayaRyzen AI Max+ PRO 495 (Radeon 8065S, 192GB, 160GB to the GPU, Windows 11)43.2Maya v1.1.37 --bench/--report, HIP: decode 43.2 (code 46.1 / 50.9, prose 32.6); prefill 453 (278 at 2K). All experts in GPU memory. GLM-5.3-Flash Maya-S-v2 31.5 on the same machineMeasured2026-11-03link
DeepSeek-V4.1-Flash-1017StrataRTX 5090 + 128GB DDR5 + i9-13900KF (Windows 11)40.66Strata Terminal 0.2.4.1: three runs 34.57 / 40.66 / 42.57, median 40.66 (higher than other reports on similar setups, but the 3 runs agree and a screenshot is attached)Measured2026-10-24link
DeepSeek-V4.1-FlashMLX2× Mac Studio M3 Ultra 512GB4040–66 tok/s decode (acceptance rate 70–81%); ~1,100 PPMeasuredlink
DeepSeek-V4.1-Flash-1017Strata1× RTX PRO 5000 72GB + Ryzen 9 9950X3D (Linux)39.05Strata Terminal 0.2.4.2, DSV41 engine 0.1.49.2, 512-token output, n=3: DSpark off 39.05, on 42.77 (API level); 240K retrieval read-in 928.5Measured2026-10-25link
DeepSeek-V4.1-Flash-1017vLLM4× Tesla V100 SXM2 32GB (SXM2-to-PCIe carrier board, NVLink) + Threadripper PRO 3945WX + 256GB DDR438.31Cat-vLLM fork (#1277), TP4, one user: 8K thinking off 38.3 (36.6–39.0), 75K 35.8, thinking on 31.9–37.9; prefill about 1,000–1,200 (4K–128K); one third of the experts in host RAM; 150–175 W per card; context 524,288Measured2026-11-02link
DeepSeek-V4.1-Flash-1017StrataThreadripper 3960X + RTX 5090 + RTX 3080 Ti + 192GB DDR4 (Windows)38About 38 on average over a 6-turn coding conversation (37.4–38.7), up to 42.4 after warm-up; about 26–30 before the patch; stock defaults 19.4, settings only 29.1; prefill up to 772.1Measured2026-10-30link
DeepSeek-V4.1-Flash-1017Strata1× RTX 5090 32GB + Xeon Gold 6330 + 384GB RAM37strata-dsv41 0.1.55: median decode about 37 (32–42, including 100K–180K prompts); cold prefill 1,612 for a 139K promptMeasured2026-11-01link
DeepSeek-V4-Flash (0731)affinityThreadripper PRO 9955WX + 128GB + 2× Radeon AI PRO R970036decode 36 tok/s (~45% experts in VRAM; rest streamed/offloaded from 51GB pinned host pool)Measured2026-10-16link
DeepSeek-V4.1-Flash-1017Strata1× RTX 3090 Ti 24GB + Threadripper 3990X + 128GB DDR435Decode about 35 (MTP/DSpark on)Measured2026-11-01link
DeepSeek-V4.1-Flash-1017Strata1× RTX 5090 32GB35Strata 0.2.10 main value 35 (prose decode 35–38, low end); history: Long-generation average about 50–51 (2.2K-token test, 256K context); prefill 1,350. The numbers come from the post text; the screenshot was not checked; Strata 0.2.10 (2026-11-03): prose decode 35–38, copy/agent 68; prefill 890–2,264 at 10–36K (x.com/lyesbahi/status/2117672208592220602)Measured2026-11-03link
DeepSeek-V4.1-Flash-1017Unspecified2× RTX 3090 + 256GB 8-channel DDR4351M context: decode about 35, prefill about 900. The post does not name the engineMeasured2026-11-03link
DeepSeek-V4.1-FlashoMLXMac Studio M3 Ultra33About 6 → 33.0 tok/s (32K 6.17→33.0; 64K 5.58→33.0)Measuredlink
DeepSeek-V4.1-Flash-1017TideLM2× DGX Spark (GB10, TP)31.8Median of 3 runs, 1,024 tokens, temp 0.6: prose 31.8 / code 49.0; first token after an 8K prompt 4.8 s. tok_s is proseMeasured2026-10-27link
DeepSeek-V4-Flash (0731)Unspecified1× DGX Spark31.5x1 31.5 tok/s, TTFT 408ms; x2 total 29.8Measured2026-08-21link
DeepSeek-V4.1-FlashStrata (Strata-DS fork)RTX 5090 Laptop 24GB + Ryzen 9 9955HX3D + 64GB RAM + 2TB NVMe31.2Short prompts from a cold start 30.0–32.0 (median of 5; code prompt 31.2); through the API on a running server 33.2–39.3; 30 minutes non-stop 37.6; with 4K/32K/128K/256K prompts decode 26.1/25.6/25.2/24.4 and prefill 244/844/904/752 tok/s; skips on: HumanEval+ 90.2% (off 91.5%), GSM8K 98.0% (off 97.5%)Measured2026-10-16link
DeepSeek-V4.1-Flash-1017Strata (Strata-DS fork)1× DGX Spark (GB10, 128GB unified memory, 4TB NVMe)30.8Real coding-agent session, 48 turns (context 10K–91K), 70,211 tokens at 30.8 tok/s average (per turn 23.6–36.4, median 28.6); single runs: decode at 8K/32K/64K/128K/250K context 24.2/27.4/27.4/27.3/22.9; time to first token for an agent turn adding ≤50 tokens median 0.57 s, a new conversation with the same 10K system prompt 0.25 s; cold prefill about 500 tok/s (250K in 8.3 min); vs official MXFP4 KL 0.074 / top-1 88.6%Measured2026-10-18link
DeepSeek-V4-Flash (V4, not 4.1)ds4 (DwarfStar)Mac Studio M4 Max 128GB30~30 tok/s write; 37K prompt prefill about 2 minMeasuredlink
DeepSeek-V4.1-Flash-1017oMLX2× Mac Studio 256GB (chip not stated)30decode about 30 (the author compares 171 on a cloud API); context and interconnect not statedMeasured2026-10-27link
DeepSeek-V4.1-Flash-1017vLLM4× DGX Spark (GB10, switchless QSFP ring)29Single stream: prose 29–43 / code 65–67; 4-stream code aggregate 115–127; 111,840-token long task average 46.5 (~80 in the code phase); 32K prefill 3,655; arx off by default costs ~3%Measured2026-10-20link
DeepSeek-V4.1-FlashUnspecified2× DGX Spark29Decode 29 tok/s prose / 41 code; prefill 2,000 tok/s; top-token agreement 92%, KLD 0.07Measured2026-10-03link
DeepSeek-V4.1-Flash-1017Project Maya1× Tesla V100-PCIE-32GB (PCIe 3.0 x16, 150W cap) + Xeon E5-2680 v4 (14 cores) + 128GB DDR4-2400 (HUANANZHI X99-QD4)23.5Project Maya v1.1.61 @c5842c5 (recorded in maya-report.txt), configured context 262144, one GPU. maya-bench.txt after --calibrate: three answers decode at 22.0 / 21.3 / 27.1, one at a time; the mean, 23.5, is the headline (greedy, thinking off, up to 256 tokens per answer). The installer-default 24.5 (issue table, no matching attachment) is a note only. Prefill at 2048 / 8192 tokens: 200 / 510 after calibration, 164 / 355 with defaults. The nvidia-smi report confirms V100-PCIE-32GB on PCIe 3.0 x16. The CPU-lane experts are copied into RAM (99.5GB). maya-report.txt and maya-bench.txt are attached, so this is measured.Measured2026-11-13link
DeepSeek-V4.1-Flash-1017gufo1× Bosgame M5 (Strix Halo, 128GB)23.3Greedy, measured over HTTP: depth 0 / 32K / 128K: prefill pp2048 1,077 / 993 / 700; decode 23.3 / 22.9 / 21.7; DSpark on mixed text 28.0 / 30.1 / 27.2; DSpark on repetitive text (separate, reads high) 65.3 / 59.6 / 55.2; aggregate (separate) 8 users 78.5; editing a source file 67.6. tok_s is decode at depth 0Measured2026-10-27link
DeepSeek-V4.1-Flash-1017Strata1× AMD Instinct MI50 32GB (gfx906, ROCm 7.1) + 62GB RAM + NVMe (about 4GB/s)22.6Decode: default 15.8; --cpu-threads 24 gives 16.7; add --experts 5 for 19.6; add DSpark (conf 0.65) for 22.6–23. Prefill on a 7.5K prompt about 170. Single runs, 200-token answers, tested up to 8KMeasured2026-10-29link
DeepSeek-V4.1-Flash-1017ds412× RTX 3090 (PCIe 4.0 x16 + 3.0 x4) + 128GB DDR4-3200 + i7-12700K + NVMe22General prose decode 22–24 tok/s; 3,000-token agent prompt 31.5 (engine benchmark, hot experts ranked on agentic text, optimistic); prefill 3,000 tokens 340; vs FP8 KL 0.462 / top-1 71.5%Measured2026-10-18link
DeepSeek-V4.1-Flash-1017Strata3× Tesla V100 32GB PCIe 3.0 + 2× Xeon E5-2696 v3 + 128GB DDR3 (Windows 10)22Strata source @7dd5cf1 (v0.2.4.2, sm_70, CUDA 12.4), layer split auto, --dspark auto: 3072-token answers 22.0, 1024-token 17.4; prefill 7K 176. Same machine on maya-ds4 v1.1.26 (Maya-S): 27.5 / 24.2 / 425Measured2026-10-25link
DeepSeek-V4.1-FlashStrataRadeon PRO W6800 32GB (gfx1030) + 192GB RAM (Windows 11)20.2Engine built from source for gfx1030, strata_bench --long: --dspark on short-chat median 20.2 (auto/off 17.1); one live long answer 24.9 (15.7 with the drafter off). Experts in 192GB RAMMeasured2026-11-04link
DeepSeek-V4.1-FlashStrataRTX PRO 4500 32GB + RTX PRO 4000 24GB, HP Z820 (2× Xeon E5-2697 v2, no AVX2; PCIe 3.0 x16)19.9Strata 0.2.13.1 / engine 0.1.59.1, AVX1 source build. tools/strata_bench.py --long, 3-run medians, thinking off, one request at a time. Headline is 6-expert chat 19.9 (baseline and fill agree); the 5-expert --speed faster path is 31.3 chat / 26.5 agent and is separate. Context 64K. Engram hot path was in RAM. 0 failed requests. Community format, not standard-v1. Tower workstation; no rack evidence.Measured2026-11-10link
DeepSeek-V4.1-Flash (REAP 256-expert pruning)vLLM2× DGX Spark19.2c=1, DSpark k=5, 32K: code 30.4 / reasoning 28.3 / chat 27.0 / prose 19.2 tok/s (counting 49.4, author warns not to use); prefill 700–2,200; 25K prompt TTFT ~22 sMeasured2026-09-12link
DeepSeek-V4.1-Flash-1017StrataRTX 4000 Ada 20GB + Xeon w7-3455 + 255GB RAM (Windows 11)15.6Decode 15.6–17.4 (two sessions, median of 8×200 tokens); prefill 237.5 at 3.8K / 506.2 at 20K; 1,283 of 15,360 experts in VRAM and the rest in RAM. tok_s is the low end of the rangeMeasured2026-10-27link
DeepSeek-V4.1-Flashllama.cppACEMAGIC Strix Halo 128GB (€2,300) + fast SSD15>15 tok/s, peak 31.7Measuredlink
DeepSeek-V4.1-FlashStrataAMD Ryzen AI Max+ 395 128GB (Strix Halo, Windows)15Code decode 15–21; prefill 214–233 at 64K contextMeasured2026-10-29link
DeepSeek-V4.1-FlashStrata1× RTX 5060 Ti 16GB (PCIe 4.0 x8) + Ryzen 9 7900X + 64GB RAM (experts/Engram streamed from Gen5 NVMe)14.8Strata 0.2.13.1 / engine 0.1.59. tools/strata_bench.py --long, 3-run medians, thinking off, one request at a time: short-chat decode 14.8 (13.6–16.9); agent 16.0. Install defaults: context 131,072, vision on, automatic RAM store, DSpark auto. Experts and Engram stream from Gen5 NVMe, unlike the RAM-resident Z820 path for the same model. 0 failed requests. Community format, not standard-v1.Measured2026-11-10link
DeepSeek-V4.1-FlashStrataMacBook Pro M5 Max (40-core GPU) 128GB14.3Through the server (greedy, thinking off) short and code answers 14.3–19.0, prefill of a 3,000-token prompt 268; default settings give only 5.0–6.5. Engine CLI: default 6.66, DS4_BUSY_CAP=2000 gives 18.05Measured2026-10-31link
DeepSeek-V4.1-Flash-1017StrataRTX 4090 Laptop 16GB + i9-13900HX + 94GiB RAM (Gentoo)13.62.25 default: chat 13.6 / agent 16.0 / code with thinking 16.2; --experts 6: 11.2–13.7; 2.00 default: 14.9–17.3. Prefill 4K / 32K: 436 / 800 (2.00: 463 / 833). Speed depends much on RAMMeasured2026-10-28link
DeepSeek-V4.1-FlashStrata2× RTX 4060 Ti 16GB (--tp 2) + Xeon W-2145 + 110GiB (WSL2)13.2Strata 0.2.13.1 (628f749), DeepSeek engine package fetched by the launcher. The author labels this a non-standard measurement (Qwen protocol fixtures; no DeepSeek standard protocol). Headline is first-pass short decode median 13.2 (12.9–13.3); it drifted to 7.4–8.4 within about 15 minutes, cause unmeasured, and that is recorded. Configured 131,072, q4_0 KV, --tp 2. results.jsonl and the engine log are in the PR.Measured2026-11-12link
DeepSeek-V4.1-Flash-1017Strata1× Intel Arc Pro B70 32GB + Xeon w5-3435X (container limit 192GiB / 24 CPUs, host 503GiB)13Strata draft PR #5907 (head 28952cee; implementation commit 730e44a8 on upstream 13eedf55), SYCL JIT build, engine 0.1.59.1. Configured 16384, 1 slot, top-6, DSpark on with confidence 0, greedy, thinking off. Headline is 1024/256 decode median 13.0 (12.3-14.5, 3 runs). 4K 12.6, 8K 12.4. Prompt read 1K/4K/8K: 197/377/601. Raw records results.jsonl, engine timings and configs are in the bench/results folder at the PR commit. https://github.com/Niko1221/Strata/blob/28952cee2c09dad4df8d60f4541a08d1dc8a3bde/bench/results/2026-11-13-community-b70-deepseek41-sycl-1k4k8k/README.mdMeasured2026-11-13link
DeepSeek-V4.1-Flash-1017StrataRadeon AI PRO R9700 32GB + RX 9070 XT 16GB13Generation 10.1–15.7 (mean 13.0); prompt processing 18.0; ELYZA 10 tasks + 5 hard tasksMeasured2026-10-28link
DeepSeek-V4.1-Flash-1017StrataGMKtec K12: Ryzen 7 H 255 + Radeon 780M iGPU (gfx1103) + 80GB DDR5-5600 (Proxmox LXC)11strata-dsv41 0.1.58 (146df3c, TheRock ROCm 7.10.0a20251121, STRATA_EXPERIMENTAL_GFX1103=1). Headline is row-7 code-prompt 150-token decode 11.0 with --dspark on + HIPK; same-table prose 300-token decode 8.5 is not the headline. Drafter off about 6.4–6.7. Prefill 7K/19K 97.8/121.5 (HIPK). Issue includes the measurement script; same-host #5639 strata_bench short chat 7.3 / agent 7.4 are listed separately.Measured2026-11-09link
DeepSeek-V4.1-FlashStrata1× AMD RX 7900 XTX 24GB10Short-chat decode 10.0 (8.0–12.3); agent follow-ups 14.3; prefill 4K 164.6 / 16K 282.3 / 32K 252.7; 1,896 experts in VRAM, the rest read over PCIeMeasured2026-10-31link
DeepSeek-V4.1-FlashStrata1× RTX 5070 12GB + i7-12700K + 64GB DDR4 (Linux, PCIe gen5)10Same box as #5435: strata_bench short-chat decode median 10.0 (9.1–10.8); mtp-bench overall server_decode 12.2 (per-prompt 10.1–12.5); drafting off. tok_s uses short-chat 10.0Measured2026-11-06link
DeepSeek-V4.1-Flash CoderStrata1× RTX 5080 16GB + 127.4GiB RAM (Windows 11; CPU not stated)9.8Strata engine ds4-0.1 / 0.1.56. Community suite, 3-run medians. On a 16 GiB card the default Faster path (5 of 6 experts) short-chat decode is 9.8 (8.9–11.2) as the headline; full-quality 6 experts is 8.4 (7.9–8.8) and is separate. Agent 10.4 / 8.4. Thinking on, temperature 0, context 131072. CPU is not named and is not inferred from the author's other machine. Speed suite only, no correctness checks.Measured2026-11-11link
DeepSeek-V4.1-FlashStrata1× RTX 5080 16GB + Ryzen 7 7800X3D + 127.4GiB RAM (Windows 11)9.4Strata 0.1.59 (d2f09647, worktree at tag v0.2.13) / DeepSeek engine ds4-0.1 (0.1.56). tools/strata_bench.py --long --think, same protocol as the same-machine Coder set comm-331. Headline is chat decode median 9.4 (9.0–10.0, n=9); agent decode 9.6 (9.2–10.9, n=15) is separate. Default 5 experts; DSpark was not loaded (the 115GB expert set does not fit the card). Configured context limit is not stated, left null. results.json is in the PR.Measured2026-11-12link
DeepSeek-V4-Flash-0731-Fractalllama-modchipAMD Ryzen 9 9955HX + 64GB DDR5, CPU-only (0 GPU)9.06Out-of-window makeup (v2.4.0 released 2026-11-04): engine-author README table. Lossless Top-6 pure decode 9.06–9.95; tok_s is the lower endpoint; Stockfish-PVS 11.5–14.7 is separate. About 58.9 GiB RSS at 256K. Not in the 841-row baseline.Measured2026-11-04link
DeepSeek-V4.1-Flash-1017StrataRTX 5090 D v2 24GB + 128GB DDR5-4800 + i7-14700KF (Windows 10)8.6Strata 0.2.4 (DSV4.1 engine 0.1.49) defaults (--dspark auto): 256-token streaming, decode short 8.6 / 1.3K 7.9 / 4.9K 10.2, tok_s is short prompt; 4.9K prefill 240; RAM store grew only to about 50GiB, author suspects it is not at full speedMeasured2026-10-24link
DeepSeek-V4.1-Flash-1017StrataRadeon AI PRO R9700 32GB (PCIe 5.0 x16) + RX 7900 XTX 24GB (PCIe 3.0 x1) + Ryzen 5 7600X + 62GB DDR5 + Kingston NV3 2TB (Ubuntu 24.04)8.29short prompt (54 tokens, 256 new tokens) decode with DSpark off (same basis as comm-079/080/082): --ssd-cache 32, two runs 8.85 / 7.72, tok_s is their median 8.29; release defaults 6.27; with DSpark on (--dspark-conf 0.65) three runs give 9.82–9.94; prompt about 2; thinking off, greedy, numbers from the engine decode linesMeasured2026-10-24link
DeepSeek-V4.1-Flash-1017llama.cppRTX 3090 24GB + i9-9900X + 128GB DDR47.27.2; thinking off, 16K context; author says the answer has a physics error and sees little practical useMeasured2026-10-21link
DeepSeek-V4.1-Flash-1017StrataRX 9060 XT 16GB (gfx1200) + Ryzen 5 5600 + 94GiB DDR4 (Linux)6.2Decode 6.2–6.6 (6 experts) / 6.8–7.5 (5 experts), limited by the 6-core CPU; prefill rises from about 48 to about 230 with prompt length; no engine error in two daysMeasured2026-10-28link
DeepSeek-V4.1-Flash-1017Strata1× RTX PRO 6000 Blackwell Workstation 96GB + Ryzen 9 9950X3D + 128GB DDR5 (Windows 11)Windows: prefill mostly 40–130 (13.5K fresh prompt 112.5→126.5, PCIe x8→x16), decode 27–61; the author reports it as an engine problem; tok_s is emptyMeasured2026-10-26link
DeepSeek-V4.1-FlashTandem-vLLM1× RTX PRO 6000 96GB + EPYC 9534124Tandem-vLLM 0.1.1 (expert-offload build of vLLM 0.31). Engine-author table: single-stream system decode with DSpark k=3 is 124, the headline; 65 with no drafter is recorded on this row, not as a second row. 4/8/32-stream totals without a drafter, 182/284/414, are not single-stream. Official GPQA 81.3% falls to 73.2% with VQ (76.8% with FP4_TOP=52) and is noted. Context is unstated, left null. Comparison table, no raw log; claim. 2026-11-14 Tandem-vLLM 0.2.0 post by the same author: same machine, one card, one user, GSM8K decode 75.5. The post does not say whether DSpark is on, so the basis differs from the headline; note only, the headline 124 does not change. Same post: one RTX PRO 6000 at 32 streams, aggregate 414. Same post: two-card tensor parallel 93.1; two RTX PRO 6000 cards at that day's lowest price of $17,496 each already exceed $30k, so it has no row. https://www.reddit.com/r/LocalLLaMA/comments/1xvjoa7/Claimed2026-11-12link
DeepSeek-V4.1-Flash-1017TensorFold4× GB10 (DGX Spark, TP4)121TensorFold TP4 v0.3.5: decode about 121; prefill 8K 2,668 / 32K 2,900. Note: on 10-25 niklaslenz measured bertholomus 4× Spark 4.0bpw at 82.8 / 80.8; the quant and version differ, so it is not a direct comparisonOfficial2026-10-24link
DeepSeek-V4.1-Flash-1017SGLang4× DGX Spark (dual RoCE links, TP4)91.3sparkDash 1.8.8, 256 new tokens, temp 0, thinking off: c1 prose 91.3 / code 142.9 / json 130.6 (median of 3), tok_s is prose; aggregate (separate) c16 prose 359.7 / code 471.6 / json 726.9; cold prefill 4K 4,798 to 128K 6,291 (repeated-token filler, reads high)Claimed2026-10-24link
DeepSeek-V4.1-Flash-1017spark-ds41f4× DGX Spark TP4, MikroTik CRS812-DDQ 4× 200GbE RoCE89.2spark-ds41f e90b1e8a + local HEAD 9b13cc2c (image spark-ds41f:1107-local; changes are sudo/memguard and a QEval harness only, no kernel changes). Runtime is 1M context, DSpark 5 drafts, TP4. Headline is the author's correction: single-stream prose 89.2 at a 32K starting context, not the 188.3 structured figure in the title; structured 184.8 (32K) / 188.3 (standard context) is separate. C8 total 886.2 is not single-stream. Method is described, no public raw log; claim. Thread https://x.com/FT_IOxCS/status/2120833708693266798 2026-11-14, same author and machine on release 1107: one coding prompt from a 32K starting context 143.8; standard short-context C1 coding 110.5; temperature 0, thinking off. Coding is a separate basis; the headline stays prose 89.2. C16 (32K) aggregate 422.8 is not a single-stream value. https://x.com/FT_IOxCS/status/2121296132630282446Claimed2026-11-12link
DeepSeek-V4.1-Flash-1017mlx-serve2× Mac Studio M3 Ultra (tensor parallel over TB5 RDMA)88mlx-serve draft PR (not merged), llmprobe: DSpark decode 88 (56 on one Mac), 116 on predictable text; plain decode 54 (41 on one Mac); prefill 1,608 (1,018 on one Mac); 4 concurrent streams 94 aggregate (62 on one Mac). Average decode / prefill over 0.5–256K: 69.5 / 1,365 on two Macs, 51.7 / 889 on one Mac (x.com/ivanfioravanti/status/2117658847544525094); 2026-11-05 ivanfioravanti on original weights with the DeepSeek Harness: mlx-server 2×M3 Ultra and TensorFold 4×Spark both average about 52 decode, max context 815K (M3) / 400K (Spark) (x.com/ivanfioravanti/status/2118222107783873005)Claimed2026-11-03link
DeepSeek-V4.1-Flash-1017Strata4× Radeon AI PRO R9700 32GB + 251GB RAM84.7code 76.6/77.0→84.7/84.7, agent 74.2→83.3 (0.2.5→0.2.6); RAM must hold all CPU-side experts with 8 GiB free; older 0.2.4 values (2026-10-24, strata-dsv41 0.1.49, IQ2_XXS 2.00bpw GGUF, https://github.com/Niko1221/Strata/releases/tag/v0.2.4 ): output about 31, prompt about 505 at 8K and about 818 at 32K. Qualification: 0.2.4 output with DSpark 55–69, not the headlineOfficial2026-10-25link
DeepSeek-V4.1-Flash-1017Project MayaAMD Strix Halo (Ryzen AI Max+ 395, 128GB) + 2× RX 7900 XTX (OCuLink)82.5Maya-XS v2: decode 82.5; prefill for a 491K prompt 1,195.5 → 1,448.2 (the sparse-attention indexer now uses RDNA 3 WMMA). Maya-S: decode 41.7, prefill at 130K 1,378Claimed2026-11-03link
DeepSeek-V4.1-Flash-1017spark-engine3× DGX Spark (TP3, ConnectX-7 RoCE)67Short-context single stream (greedy): prose 67–70, code 114, structured 129; prose 67→125 for 1→4 streams; prefill about 5.6K at 32K–64K; first token after 229 s for a 1M prompt (about 4.5K), decode 53–56 at 256K–1M. The author also reports prose 68 for Qwen3.8-Flash-Next on 1× SparkClaimed2026-10-31link
DeepSeek-V4.1-Flash-1017mlx-serveMac Studio M5 Ultra 512GB65.7mlx-serve 26.11.2 defaults (MTP and prompt lookup on). 128K-token prompt, output decode median of 5 is 65.7 (range 62.2–70.0). Same box: prompt lookup off 65.3, drafting off 49.9; prefill 3,632, first token 35.3 s. Short requests with thinking off: prose 68.8 / code 81.2 / counting 116.9. Sixteen users with fresh 2K prompts and 512-token answers total 112.3 (one user 59.1 under that shape), not single-stream. Configured context limit is not stated, left null. No raw log; claim. Same-window HeyBohu short request 60.6 (mlx-serve, official weights, M5 Ultra 512GB) is a short run and does not change the headline. This row already states mlx-serve 26.11.2, so it is not split for that version. An ApacheAE post says an 8-bit output head changes speed; that is a note only. https://x.com/HeyBohu/status/2120662512660259266 https://x.com/ApacheAE/status/2120514025239064840Claimed2026-11-10link
DeepSeek-V4.1-Flash-1017BasaltRTX 5090 32GB (400W) + RTX 5060 Ti 16GB (180W) + 128GB DDR563.55090 + 5060 Ti each computing experts: decode 63.5 / prefill 610 (6.2K prompt); 5090 only: 53.6 / 610 warm, 595 cold; same machine llama.cpp PR #30817 --fit about 20; the author has not published the experimental branchClaimed2026-10-21link
DeepSeek-V4.1-Flash-1017sushiMac Studio M5 Ultra 96GB62.2Sushi 1.5.1. Configured 1M context. Cold ladder, 64 new tokens per rung, 3-run medians: headline is 4k decode 62.2; 64k 41.6 and 128k 29.9 are separate; about 23 at 1M. Measured 10 Nov while the engine was also serving this chat; the author calls it slightly conservative. No raw logs, so this is a claim.Claimed2026-11-10link
DeepSeek-V4.1-Flash ISTA-DAS Coder (192 of 384 experts per layer)Strata4× Radeon AI PRO R9700 32GB (Linux)59.2Decode: 4× R9700 59.2, 1× R9700 22.8, Strix Halo 20.9–21.9, M5 Max 19.1–22.5, RTX PRO 6000 Blackwell about 100. One R9700 with RAM limited: 40 GB 20.9 / 32 GB 18.5 / 24 GB 15.0 / 16 GB 8.0. Model card: SWE-bench Verified 77.00 (base model 81.20). Model huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUFOfficial2026-11-03link
DeepSeek-V4.1-Flash-1017Strata1× RTX 5090 (inference only) + RTX 3070 Ti (display) + 128GB DDR5 + i9-13900KF (Windows 11)53.8Strata 0.2.11: median 53.8 (mean 52.0, min 41.5 max 57.8). Screenshot attached. tok_s uses the median 53.8Claimed2026-11-06link
DeepSeek-V4-Flash (V4, not 4.1)UnspecifiedStrix Halo + R970051.1vendor homepage 86 tok/s; independent retest 51–57 (median 51.1)Claimedlink
DeepSeek-V4.1-Flash-1017FastLocalExperts1× RTX PRO 4500 Blackwell 32GB + Ryzen 9 9950X3D + 120GB DDR5 (Linux)50.6Engine-author README: v1.2.5, three fresh starts 49.8/50.8/50.6, headline 50.6 (temperature 0, lossless draft); no-draft 37.6 is separate. Hardware is 9950X3D + 120GB + PRO 4500. The README states that the speed figures above use a 262,144-token context shared by two slots. The same author's 11-11 Reddit post of 49.9 (3-run median, no-draft 37.2) does not name a version and does not replace this headline or get its own row. The README's QFN FLE-IQ3_M 91.2 does not name a version and is not ingested. Author table without per-run raw logs, so this is a claim.Claimed2026-11-11link
DeepSeek-V4.1-Flash-1017oMLXMac Studio M5 Ultra 512GB (same box as x-296)49.1oMLX 0.7.4.dev1, same box and 128K-prompt protocol as x-296: output decode 49.1. Prefill 3,848, first token 33.3 s. Short requests with thinking off: prose 51.5 / code 68.9 / counting 90.2. Sixteen-user total 110.2 (queues past 8) is not single-stream. Configured context limit is not stated, left null. No raw log; claim.Claimed2026-11-10link
DeepSeek-V4.1-Flash-1017Project MayaRTX 4090 24GB + 192GB RAM44.1Decode 44.1; prefill 539 at 8KClaimed2026-11-04link
DeepSeek-V4.1-Flash-1017delta-inference-engineRTX 3090 24GB + i5-13600K + 192GB DDR5-5600 (Windows)41.7Delta Inference Engine v0.93 (same author, same 3090 + 192GB box): UD-IQ2_XXS at 192GB write speed 41 (flat vs prior); same table at 128/96/64GB is 34.0 / 29.3 / 21.0; UD-IQ1_S at 192/64GB 41.0 / 30.6; prefill up to about 740 at 64K (192GB); 57/99 on the 99-question set (prior ds4 Q2 was 52). Note: Unsloth Dynamic quant file revisions affect backends differently and are not cross-backend comparable (see Strata #5296 gfx1151 prefill regression vs this row's decode gain). tok_s uses UD-IQ2_XXS@192GB = 41. On 2026-11-09 the engine author musicurgy shipped v0.94: same 3090 + 192GB box, UD-IQ2_XXS greedy write speed 41.7 (default sampling 41.3); long-prompt prefill at about 60K is 811 (747 the same evening on 0.93). An extra RTX 3060 in an x4 slot reaches 42.9 / 42.2. README and Reddit tables match. Headline updated to the 0.94 greedy 41.7; level stays a claim.Claimed2026-11-05link
DeepSeek-V4.1-Flash-1017dsx2× RTX 3090 24GB + i7-12700K + 128GB RAM41dsx (Rust + CUDA), expert deferral (--defer-rank 3, changes the computation) + FP8 projections: free generation 41Official2026-10-25link
DeepSeek-V4.1-Flash-streaming-repack-exl3-3.0bpwmlx-serveMacBook M5 Max 128GB39.939.9 tok/s decode; 81.3 tok/s prefill; 256K prompt decode 27.1, prefill 456.3, first token 574.5 s; about 1M (1,047,488) prompt decode 11.1, prefill 262.6, first token 3,988.9 sClaimed2026-10-06link
DeepSeek-V4.1-Flash-1017Project MayaRTX 3090 + RTX 3080 Ti + 128GB RAM38After a Maya update, decode 27 → 38 (with the option that skips low-weight RAM experts, lossy); prefill at 74K 1,225 → 1,840Claimed2026-11-03link
DeepSeek-V4.1-Flash-1017libosmosAMD Strix Halo (Ryzen AI Max+ 395, 128GB)38Engine author: up to 38 with drafting, prefill >450 (Reddit post says >480). tok_s uses 38. 2026-11-09 libosmos v0.5.9 on the same Strix Halo 128GB: ud-iq1_s plain/drafted 19.5/42.5; ud-iq1_m 17.5/35.4; ud-iq2_xxs 14.4/26.5. Context omitted. Drafted 42.5 is above this row's peak 38, but without context the headline stays 38.Claimed2026-11-05link
DeepSeek-V4.1-Flash-1017Unspecified2× AMD Strix Halo (USB4 link)35.3Decode 35.3; 41.9–70.6 with their own speculative drafter; prefill about 1,100 (8K–45K); aggregate (separate) 8 requests about 114Claimed2026-11-02link
DeepSeek-V4.1-FlashFront_Eagle739 streaming stackMac Studio M3 Ultra 512GB + RTX 5090 PC (32GB RAM, RAID0 NVMe)35Prefill on long prompts about 3,800; decode 35 (prose) to 60 (code); Mac only about 900 / 60Claimed2026-10-30link
DeepSeek-V4.1-Flash-1017vLLM1× DGX Spark (cold experts on the SSD)30.3Based on the ViC305 vLLM runtime, vision on: decode 30.3 on real 17–21K-token Hermes requests; 35–40 on short prompts; prefill about 910–1,020 at 32K–128K; ViC305 relayed AgentSparko claiming decode 40 and prefill 1,020 (x.com/ViC305/status/2118028345615229131; relayed, main value unchanged)Claimed2026-11-04link
DeepSeek-V4.1-FlashProject Maya2× RTX 3090 24GB + 1× RTX 2070 Super (DSpark drafter only; CPU and RAM not stated)30Project Maya v1.1.59. Two RTX 3090 cards run the main model; an RTX 2070 Super runs only the DSpark drafter; the n-gram table is on SSD. Decode 30. The author names the new Unsloth v3.1 q2 quant; the exact level is not stated. Context limit, CPU and RAM are not stated. No raw logs, so this is a claim.Claimed2026-11-13link
DeepSeek-V4.1-Flash-1017Strata1× Radeon AI PRO R9700 32GB + 251GB RAM29.62.25 code 26.1→29.6, agent 23.1→27.0; 2.00 code 24.5→27.6, agent 21.9→27.1 (0.2.5→0.2.6); load of 2.00 about 55 s (was 34 s); older 0.2.4 value (2026-10-24, strata-dsv41 0.1.49, IQ2_XS 2.25bpw GGUF, host RAM not stated, https://github.com/Niko1221/Strata/releases/tag/v0.2.4 ): output about 11Official2026-10-25link
DeepSeek-V4.1-FlashXolotl1× RTX 4090 + 90GB RAM27Up to about 27; experts are tiered across VRAM, pinned RAM and NVMeClaimed2026-10-31link
DeepSeek-V4.1-Flash-1017Tandem-vLLM2× RTX A6000 48GB (main card + helper GPU; no P2P or NVLink needed) + EPYC 7F32 (8 cores)26.7Tandem-vLLM 0.2.0 (expert-offload build of vLLM 0.31), model package ewhacc/DeepSeek-V4.1-Flash-1017-Tandem-VQ2. Engine-author table: one-user GSM8K decode. Headline is helper-GPU mode, 26.7. Same machine with one card: 22.1; tensor parallel: 11.8 (the two cards split the 8-core CPU). A6000s limited to 24GB to copy the 3090/4090 memory shape: one card 16.8 / TP 12.5 / helper 21.7; these simulated results are notes only. Context is not stated, left null. No raw logs, so this is a claim. The same author's RTX PRO 6000 row is comm-371. The post does not say whether DSpark is on.Claimed2026-11-14link
DeepSeek-V4.1-FlashProject Maya2× RTX 3090 + 128GB RAM26Same-box head-to-head: Maya 1.1.54 median decode 26.04; Strata 0.2.11 median 22.75 (Strata read/total time about +19–22%). tok_s uses Maya 26.0Claimed2026-11-05link
DeepSeek-V4.1-Flash-1017Tandem-vLLM1× RTX 6000 Ada 48GB + EPYC 7F32 (8 cores)25.5Tandem-vLLM 0.2.0 (expert-offload build of vLLM 0.31), model package ewhacc/DeepSeek-V4.1-Flash-1017-Tandem-VQ2. Engine-author table: one card, one user, GSM8K decode 25.5. The 2x A6000 row from the same post is comm-414. Context is not stated, left null. No raw logs, so this is a claim. The post does not say whether DSpark is on.Claimed2026-11-14link
DeepSeek-V4.1-Flash-1017StrataRTX 5090 32GB + R9 9950X + 96GB DDR525Strata 0.2.4, 512K context: decode 25–40 (the lower bound, 25, is used; the author is still tuning); prefill not measuredClaimed2026-10-24link
DeepSeek-V4.1-FlashProject Maya2× Tesla V100 32GB (150W limit) + i9-10940X + 192GB RAM + SATA SSD24.83 Japanese short stories, decode 22.8 / 24.8 / 25.0 (first run 17.3, faster after the cache warms up); thinking off, 32K context; GPUs stay below the power limit, the 14 CPU cores are sometimes at full loadClaimed2026-10-21link
DeepSeek-V4.1-Flash-1017 CoderStrataRTX 3090 + 64GB DDR423.2Strata 0.2.10: decode 23.2, prefill 659 (older version, same quant: 6–10 / about 10)Claimed2026-11-04link
DeepSeek-V4.1-Flash-1017Unspecified1× RTX 5070 Ti 16GB + Ryzen 5 7600 + 64GB DDR5 + 2× NVMe (Windows 11)20.3Experts stream from NVMe: decode 20.3 at short context (17–20 in real agent turns), fast preset 23–24.7; 16–20 at 256K / 512K; agent-turn read about 418. On the same PC the original fork about 10, Project Maya v1.1.21 3.9. 2026-11-04–06 og build: server short decode 18.4–18.8 (teacher-forced 21.3), 15–19 with 256K–1M already filled; 1M prefill 888; same-host Maya v1.1.54 about 3.3. Same ballpark as this row's ~20.3; keep headline.Claimed2026-11-02link
DeepSeek-V4.1-FlashTandem-vLLM1× RTX 5090 32GB + i5-14600KF + 128GB DDR4-320017Same post, engine and quant as comm-371, hardware is a 5090 + i5-14600KF + 128GB DDR4-3200. The author says it runs, slowly: one user 17. Four-stream total 20 is not single-stream. The drafter switch is not stated and is not inferred. Context is unstated, left null. Claim.Claimed2026-11-12link
DeepSeek-V4.1-FlashXeMayaIntel Arc Pro B70 32GB + Arc Pro B65 32GB14.56The engine-author README explicitly names the published 0.1.0-alpha.1 build, B70+B65, regular (not Coder) IQ2_XXS. Headline is decode 14.56 (prefill 11.4). The 576-token prompt and 1,024-token decode are load lengths, not a configured context limit, so context is left null. Speed only, no correctness check. The same table's GLM 17.4/20.0 and QFN 76.6/57.1 rows do not name a build and are not ingested or pinned to alpha.1 or alpha.2. Author README table without raw logs, so this is a claim.Claimed2026-11-10link
DeepSeek-V4.1-Flash-1017vLLM2× DGX Spark (GB10, direct QSFP/RoCE; experts streamed from local NVMe)14Recipe-author README. Mia two-Spark kit at 6f7d1590, vLLM, DSpark, official 1017 native weights. Measured 2026-11-06, greedy, thinking off, bench.py decode. After a warm cache, prose is 14.0–14.6; the headline takes the lower 14.0. Code 20.7–21.8 is separate. Four streams total 23.5 and is not single-stream. Context 262K. The table is an author summary without per-run raw logs, so this is a claim. Patch hash 8047fc is not stated in the README and is not recorded.Claimed2026-11-06link
DeepSeek-V4.1-Flash-1017llama.cppMinisforum MS-S1 MAX (Ryzen AI Max+ 395, 128GB) + Radeon AI PRO R9700 32GB (USB4v2 eGPU dock)14generation about 14 / prompt processing about 110–130; 128K context; load about 3.5 min; R9700 holds the first 10 layers and the output layer, the iGPU holds layers 10–39 and the context, the CPU computes experts of 4 layers; GPU-to-GPU transfers go through a pinned host bufferClaimed2026-10-22link
DeepSeek-V4.1-Flash-1017vLLMRTX 5070 Ti 16GB + 128GB RAM + NVMe11.3211.32 after tuning (7.18 before, +58%); author says it is still experimentalClaimed2026-10-21link
DeepSeek-V4-Flash (V4, not 4.1)ds4 (DwarfStar)1× RTX 3090 + 64GB RAM10.18Final configuration (set 7), steady-state decode, grand mean over four prompt types 10.18 ± 1.17 (code 10.11 / math 9.95 / prose 11.74 / synthetic 8.91), N=20 warm runs per set; baseline 1.02. With a 7,000+ token context under ds4-agent load it is 5.2–5.5. Output equals the baseline bit for bit (max_abs_diff 0). tok_s is the grand meanClaimed2026-10-27link
DeepSeek-V4.1-Flash-1017StrataRTX 3060 12GB + 46GB RAM + NVMe (Linux)9.9output about 8–10 (code 9.9, agent 7.8); prompt 149 at 8K; DSpark off by default, on adds 2–13%; hot experts in RAM, the rest read directly from SSD. The "Known limits" section of the same release notes gives 4–8 for 12GB cards, which does not agree with the speed table; tok_s is the speed-table code value 9.9Official2026-10-24link
DeepSeek-V4.1-FlashFront_Eagle739 streaming stack1× RTX 5090 + 32GB DDR5 + RAID0 NVMe (about 25GB/s)7Normal decode 7; about 20 when the router prefers experts in VRAM (this changes the output); prefill up to 1,770Claimed2026-10-28link
DeepSeek-V4.1-Flash-1017Strata1× RX 9070 XT 16GB + 5800X3D + 64GB DDR4 + PCIe 3.0 NVMe71M context, decode about 7; with the KV splice patch, the first token of a new agent session comes in about 5 s instead of about 100 s (system prompt about 37K tokens)Claimed2026-11-02link
DeepSeek-V4.1-FlashUnspecifiedsingle RTX 5090 32GB + 126GB RAM + NVMe5.1cold 5.1 tok/s; 21.4 after cache warmRelayedlink
DeepSeek-V4.1-Flash-1017UnspecifiedRTX 4070 12GB + Ryzen 7 5800X3D + 128GB RAM5about 5 (derivative weights, engine not stated)Claimed2026-10-25link
DeepSeek-V4.1-Flash-1017StrataRTX 5070 12GB + 64GB RAM + Ryzen 5 7600 + NVMe (Windows)4.64output about 5 (code; 4.64 with DSpark off); prompt 231 at 8K; RAM expert store 4.3–4.9 against 1.8–2.2 with mmap; CPU-bound, DSpark gives no gainOfficial2026-10-24link
DeepSeek-V4.1-FlashInferSSDMac Studio M3 Ultra 96GB + a pair of OWC Envoy Ultra 2TB TB5 SSDs (RAID 0 across ports, one drive per port)4.4InferSSD (version not stated) streams expert weights from SSD; the 473 GiB Q8_0 pack does not fit in the 96GB memory. RAID 0 across two TB5 SSDs decodes at 4.4–4.6, versus 3.5–3.7 on a single Envoy (about the same as the internal SSD), a gain of about +23–26%. Prefill about +67–77%; a 4,000-token prompt starts answering in about 2 minutes instead of 3.5. tok_s uses 4.4. The author wrote 763B; an 11.15GB/s figure in the comments is not the author's own measurement and is not usedClaimed2026-11-07link
DeepSeek-V4.1-Flash EXL3 3bpwvLLM2× DGX Spark (headless mode, unlocks +2GB)author says 25–40 tok/s (hard prose→structured), multi-stream >50; explicitly makes no precise claims (no single value, tok_s left blank); KV pool about 3M tokens (shared by up to 6 concurrent)Claimed2026-09-18link
DeepSeek-V4.1-Flash (510GB weights)Unspecified1× DGX SparkHTML game 36.6; SQL 31.8; Python 24.3; long story 17.1; PP 337 (degrades beyond 2,000 tokens)Relayedlink
DeepSeek-V4.1-FlashUnspecified3× DGX SparkProse 37.9; 4 streams total 78.6 (independent SGLang C1 26.5)Relayedlink
DeepSeek-V4.1-FlashTensorFold4× DGX SparkMXFP4: C1 peak 77.2, 6 streams 214; TensorFold: prose 87, code 124.8, 4 streams 246.8; PP 5,800–5,925Relayedlink
DeepSeek-V4.1-FlashUnspecified4× DGX Spark (direct-connected, no switch)up to 494 tok/s total at 32 requestsRelayedlink
DeepSeek-V4.1-FlashMLXM3 Ultra 512GB / M5 Max 128GB~14.6 / 15 (SSD streaming), 25 (RDMA tensor parallel)Relayedlink
DeepSeek-V4-Flash (V4, not 4.1)oMLXMac Studio M3 Ultra 512GB35 (oMLX, Viticci); ds4 short 35.5 / long 26.6; PP 448.8Relayedlink

Numbers come from each source. This project did not measure them again. Models and settings differ between rows. Source levels: Official = project or vendor; Measured = the poster ran it (an engine or recipe author must also publish raw data and reproduction steps); Relayed = second-hand; Claimed = prediction or no reproduction details.