⚡ Local Inference Recipes

English | 简体中文 | 日本語

Setups with NVIDIA DGX Spark: local inference speed (tok/s), sourced

As of 2026-11-14, this dataset has 157 recipes for NVIDIA DGX Spark. 109 were measured by the poster. Fastest measured single-device, single-stream text result: 130.1 tok/s with Qwen3.8-Flash-Next on TensorFold (1× DGX Spark; Measured, 2026-10-28; source). 132 entries have a date, from 2026-07-29 to 2026-11-14.

Filter these in the explorer →

ModelEngineHardwaretok/sMetricSource levelDateSource
Qwen3.8-27BTensorFold2× DGX Spark (tensor parallel across both)190Decode about 190–200 on code, about 70 on prose; prefill at 2K about 3,250 (was about 1,100); KL against BF16 about 0.03–0.04. tok_s is the lower code valueMeasured2026-11-02link
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
Qwen3.8-Flash-NextTensorFold1× DGX Spark130.1New Mia recipe (NVIDIA FP4 QAD default): code single-stream peak 130.8 (mean of 3 is 130.1), thinking off; MTP drafts off about 25.5; aggregate (separate) 2 streams 186.6 / 4 streams 288.1Measured2026-10-28link
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
Ornith-1.0-35BvLLM1× DGX Spark109Warm single stream, greedy: code about 109, math about 116, prose about 59Measured2026-11-01link
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
GLM-5.3-Flash EXL3TensorFold4× DGX Spark (works with or without switch, TP4 branch)951 stream 95 tok/s (prose); 4 streams 163Measured2026-10-10link
Qwen3.6-35B-A3BvLLM1× DGX Spark95single session 95 tok/s; 8 sessions 317 tok/sMeasured2026-07-29link
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
GLM-5.3-FlashTensorFold4× DGX Spark (GB10, switchless ring TP4, GPU clock cap 2,200 MHz)93.3Vince v2.4 overlay (TensorFold v0.6.0 Python engine, 30 files changed; not Mia's default Zig engine). Hardware, model and quant match x-018 (Mia's original 4×Spark EXL3 prose 95), but the engine does not, so this is a new row. sparkDash 2.70.0, 2026-11-11, temperature 0.7, thinking on, one run: headline is single-stream prose decode 93.3; code 138.1 is separate. Two-stream prose total 132.6 / eight-stream total 223.8 are not single-stream. Served window 1,048,576. The same README's earlier v2.1 figures (sparkDash temperature 0, thinking off: prose 90.4; own protocol 72.9) are separate. Recipe-author table with raw JSON and reproduction steps, so this is measured (one run). The recipe's gain is in prefill, not decode.Measured2026-11-11link
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
GLM-5.3-FlashUnspecified3× DGX Spark + 1× Acer GN100 (200G ring, no switch)88.1One chat 88.1; aggregate (separate) 8 chats 177.2; a 262K prompt finishes in 99 sMeasured2026-10-29link
Ornith-1.5-122BTandemLLM1× DGX Spark85.3plain about 38; MTP + lookup: prose 85.3 / code 114.2 / copy 124.8; drafted output is token-identical to greedyMeasured2026-10-21link
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
GLM-5.3-Flash (320B/18B active)TensorFold2× DGX Spark83One user, EXL3 against NVIDIA FP4-Local (TensorFold 1.1.9): JSON 82 / 52, code 83 / 47, rewrite 73 / 52, explain 58 / 42, story 46 / 36; 47 at 128K context; aggregate (separate) 4 users 82 / 56. Thinking off: GPQA Diamond 69% / 71%, IFEval 81% / 64%Measured2026-11-02link
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
Ling-3.0-flashSGLang2× DGX Spark82JSON 82, code 77; 0 of 8 pelican SVGs brokenMeasured2026-10-30link
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
Nemotron 3.5 Bolt 74B-A9Bllama.cpp1× DGX Spark80.4greedy decode mean 80.4 tok/s (math 108.2 / code 81.2 / summary 69.6 / prose 62.6; 4 tasks); sampling 74.1; 262K context; ~70 tok/s still at 131K prompt; 38.3 without speculationMeasured2026-10-17link
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-V4.1-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-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
Nemotron 3.5 Bolt 74B-A9BTensorFold1× DGX Spark 128GB72.5Mia recipe: decode prose 72.5 / code 63.4; 500 words in about 10.2 s; first token at 128K 67.8 s. Median of 2 runs, 512 tokens, thinking offMeasured2026-11-03link
Qwen3.8-Flash-NextExLlama1× DGX Spark71.7code 102.6; prose 71.7; JSON 66.2 tok/sMeasuredlink
Nemotron-3-Nano-30B-A3BOllama1× DGX Spark7171 tok/s (fastest in 30B-class comparison; QwQ on Ollama 10 tok/s)Measured2026-10-14link
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-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
KAT-Coder V2.5 35BvLLM1× DGX Spark64.6Decode code / prose 64.6; C1 62.4; aggregate (separate) C2 105.5 / C4 131.6 / C8 130.1; 4 slotsMeasured2026-10-28link
Qwen3.8-Flash-NextTensorFold1× DGX Spark64.41 stream 64.4 tok/s (prose); 8 streams 200.9; 32k prefill 2,630 tok/s; @yume_arasaki roundup (Mia data + reproduction): single-stream prose 63.6, 4 streams total 114.5 (x.com/yume_arasaki/status/2106837882870886438)Measured2026-10-08link
Qwen3.8-27BUnspecified1× DGX Spark (GB10)64reproduces MiaAI_lab recipe: 1/2/4/8 streams 64.0/104.3/155.1/238.6 (Mia: 58.3/108.5/135.1/220.6); the @yume_arasaki roundup also lists Mia code 136.9 (x.com/yume_arasaki/status/2106837882870886438)Measured2026-10-01link
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
GLM-5.3-Flash EXL3Unspecified2× DGX Spark (GB10)61.43reproduces MiaAI_lab recipe: 1 stream 61.43 (Mia 60.4); 4 streams 107.8 (Mia 108.8)Measured2026-10-03link
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
MiniMax-M3.1-FlashSGLang4× DGX Spark (TP4, switchless RoCE ring)59SGLang TP4 / nnodes=4, NCCL 2.30.7 Ring, fp8 KV: single-stream decode 59.0 (TTFT 0.183 s); prefill 32K 1,121 / 266K 1,248; aggregate (separate) 6 concurrent 122.9. Recipe from the knapcio SGLang profileMeasured2026-10-25link
gpt-oss-120bllama.cpp1× DGX Spark59gpt-oss-120b 59; Qwen3-8B 44; Qwen3.8-27B dense 13; mean of 3 runs. Efficiency 1.15 tokens/W, 0.84 for the 8BMeasured2026-10-29link
Qwen3.8-Flash-NextExLlama1× DGX Spark (GB10)58.4code 58.4, prose 53.9; aggregate (separate) 8 requests 41.5 (1 slot, queued); 32K prompt read 662; Mia TensorFold recipe on the same machine code 43.2Measured2026-10-26link
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
Qwen3.8-Flash-NextvLLM2× DGX Spark54.41 stream 54.4 tok/s, 8 streams 207, 32k prefill 2,962 tok/s (all mia-ai.net; the post says about 50 single-stream / about 212 at 8 streams)Measured2026-08-31link
GLM-5.3-FlashTensorFold2× DGX Spark (TP2)541 stream 54 tok/s; 2 streams 83; 4 streams total 142 (old 0.3.4: 40/53/54)Measured2026-10-01link
Qwen3.8-Flash-NextNVIDIA Flare1× DGX Spark52.7Recipe v2 (repo sxuff/qwen38-flash-next-flare, commit d37c2628e0; results/v2 has raw data and all pins). This recipe, A: Flare 0.1.0 @ca1f960 + 4 patches, weights nvidia/Qwen3.8-Flash-Next-FP4-Local @b9ce18a6, MTP 3 drafts, greedy. Frozen flashnext-tf405 test: 4 prompts × 3 repeats × 400 tokens, 3 fresh engine starts per config. Headline is one-request decode 52.7 (52.0-53.3). Same test on Hyperflare (NVIDIA/flare-local-inference renamed) @8a887202: MTP 51.3, DFlash2 55.5. Eight coding agents aggregate: A 85.4, Hyperflare MTP 88.3, not single-stream values. Measured 2026-11-12 to 11-13. Old headline: X post of 2026-10-26, single stream about 50 https://x.com/filicroval/status/2114454660396564597 . New post https://x.com/filicroval/status/2121282624471605457 . Earlier post (2026-10-24): FP4 single stream 50.7; TensorFold EXL3 on the same machine 75.7; 32K prompt read in 14.1 s; 8 agents aggregate 77.6, TTFT 7 s, not a single-stream value. Post https://x.com/filicroval/status/2113985358320840881Measured2026-11-13link
Ornith-1.5-122BTensorFold1× DGX Spark52.6thinking off: single-stream prose 52.6 / code 83.3; 2-stream aggregate prose 68.0 / code 103.4; 4-stream aggregate prose 90.4 / code 111.0; prefill 4K 2,450 / 32K 1,955 / 128K 1,230; about 69 GiB on device (with DFlash2)Measured2026-10-21link
Qwen3.8-Flash-Next-hibrid48 (uncensored)vLLM2× DGX Spark (TP=2)51prose 51 tok/sMeasured2026-10-03link
Nemotron 3.5 Super VL 120Bpulsar-inference-stack1× DGX Spark511 concurrent 51; aggregate (separate) 8 concurrent 225. 2× Spark (TP2, RoCE): 78 → 307, p50 first token at 8 concurrent falls from 1.45 s to 0.67 s. Both passed a 1-hour test at 8 concurrent. The FlashInfer auto-tuning at first start causes an OOM on one Spark; set MAX_JOBS=2 to prevent itMeasured2026-10-31link
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
Qwen3.8-Flash-NextUnspecified1× DGX Spark (GB10)50.3reproduces MiaAI_lab recipe, total per stream count: 1/2/4/8 streams 50.3/75.8/117.5/162.0 (Mia: 48.7/74.6/113.7/…)Measured2026-09-19link
Qwen3.8-Flash-NextvLLM1× DGX Spark48.71 stream 48.7 tok/s (mia-ai.net, not in the post); 8 streams 117 prose / 180 code (post); peak memory 91 GiBMeasured2026-09-19link
DeepSeek-V4.1-FlashvLLM2× DGX Spark48.535K agent prompt 48.5 tok/s; ≈30–32.6 at 17K–559K depthMeasuredlink
GLM-5.3-Flash (320B/18B active)TensorFold2× DGX Spark47.1One user, real prompts (TensorFold / vLLM): JSON 51.5 / 50.3, code 47.1 / 41.0, rewrite 52.0 / 40.7, explain 42.0 / 33.3, story 36.3 / 24.6; equal at 4 users; a 128K prompt is read in 91 s, turn 2 comes from the cache in 1.3 s. tok_s is the code valueMeasured2026-11-01link
Qwen3.8-Flash-NextvLLM1× DGX Spark46.5median of 5, temp 0.7, 300 tokens: JSON 62.4 / code 58.0 / explain 46.5 / rewrite 44.2 / story 33.6; random text 21.3Measured2026-10-23link
MiMo-V2.6-Flash-MOPDTensorFold2× DGX Spark (RoCE)46.15single-stream code 63.4 / prose 46.15 / structured 53.1; 4-stream aggregate: structured 103.6 / code 81.4 / prose 78.3; cold prefill 1,857–1,941; 2.9M FP8 KV poolMeasured2026-10-21link
Qwen3.8-27BTensorFold1× DGX Spark45.6TensorFold 1.2.2 (niklaslenz, 2026-11-05): code 49.0 / prose 45.6; 2K prefill 889 (was 354 on 1.2.1), 128K prefill 660; 8 streams 108.8 aggregate (x.com/niklaslenz_ai/status/2118316147686654429). History: ashxhart earlier TensorFold prose was 58.3 single-stream / 108.5 at 2 streams / 220 at 8 streams (https://x.com/ashxhart/status/2105745818137280545). tok_s uses 1.2.2 prose 45.6Measured2026-11-05link
MiniMax-M3.1-FlashNVIDIA Flare4× DGX Spark45C1 45; aggregate (separate) C2 61 / C4 75. The author says Flare is not fully optimized yetMeasured2026-10-27link
Qwen3.8-27BvLLM1× DGX Spark42.2spark-arena leaderboard llama-benchy (128 output): 1 user 42.2 tok/s, prefill 2,158; 100K 29.5; 16K 2-user aggregate 60.6 | author's own: 25-prompt median 65.1 (15.35 ms/token); quality vs BF16 99.3–101.2% recoveryMeasured2026-10-17link
Qwen3.8-27BSGLang1× DGX Spark42single stream 42 tok/s; C8 114; TTFT 0.2s; 1.86M KVMeasured2026-08-23link
Nemotron 3.5 Super VL 120BTensorFold1× DGX Spark (GB10)40.6code 40.6, prose 42.4; first token after a 32K prompt 14.3 s; aggregate (separate) 8 requests 84.2Measured2026-10-25link
GLM-5.3-FlashTensorFold2× DGX Spark38.7MiaAI 1104 recipe (TensorFold, 1M context, 8 slots, GPU clock capped at 2200 MHz). niklaslenz: prose 38.7 / code 42.4 (thinking off, 512 tokens, median of 2); with Mia's own harness he matches 47.8 (thinking on), 4 streams 87.7 aggregate, 128K prefill 1,787. jmurillocode: prose single-stream thinking off 33.7 / on 40.3, 8 streams 89.9 / 101.1, prefill about 1,747 (x.com/jmurillocode/status/2118322418523480496). JASONMCNAB: within 5% of Mia's published values, 1 stream 50.3 vs 47.8 (x.com/JASONMCNAB/status/2118312375732367487). Mia's own post reports about 48 / about 87 at 4 streams (x.com/MiaAI_lab/status/2118005417494155664). History: the author's earlier EXL3 recipe was about 60.4 single-stream / 108.8 at 4 streams (x.com/MiaAI_lab/status/2105598458547646875). tok_s uses niklaslenz prose 38.7Measured2026-11-05link
Qwen3.8-27BSGLang1× DGX Spark36.59HumanEval+ 50 problems, thinking on, c=1 net decode: plain 12.65 / MTP 24.2 / DSpark 27.96 / DFlash2 36.59 tok/s (tok_s uses DFlash2 code-task net decode; post has no prose breakdown)Measuredlink
Qwen3.8-Flash-NextTensorFold1× DGX Spark (GB10)36.24-engine agent test, 30K context, 3 runs each: TensorFold + Mia patches 36.2, NVIDIA Flare 31.2, stock TensorFold 30.2, vLLM 26.4; JSON output 67–70, prose 31–34Measured2026-10-25link
GLM-5.3 (755B full, not Flash)TensorFold4× DGX Spark (GB10, RoCE v2, TP=4)35.3Single-request decode after a 128K prompt: MTP 35.3 / DFlash2 42.9; 256K 33.2 / 40.7; 512K 32.2 / 37.8; first token after 120.1 s at 128K (MTP); coding prompt, low reasoning effort, 4-bit KV, GPU clock limited to 2.1 GHz, no prefix cache. tok_s is the default MTP value at 128KMeasured2026-10-27link
Qwen3.8-27BvLLM1× DGX Spark34.89native stack 34.89 tok/s (MTP depth 3); same GGUF + llama.cpp without MTP: 11.17Measured2026-09-26link
GLM-5.3-FlashTensorFold2× DGX Spark34stock TensorFold: single-stream 34 / 8-stream aggregate 59; 1M context; cold start about 2.5 min (author's prior vLLM NVFP4 recipe about 16 min / single-stream 28). GitHub recipe repo with ./run.sh. Same cards as x-017 (MiaAI 1104, prose 38.7) but a different recipe. tok_s uses single-stream 34Measured2026-11-05link
Qwen3.8-27BvLLM1× DGX Spark32.01HumanEval+ 50 problems, thinking on, c=1 net decode: plain 10.95 / MTP 20.89 / DSpark 25.58 / DFlash2 32.01 tok/s (tok_s uses DFlash2 code-task net decode; the post has no prose breakdown)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
GLM-5.3-FlashvLLM2× DGX Spark31.21 user 31.2; aggregate (separate) 2/4 users 45.3/54.7; 8k prompt read in 3.6 s. 2026-11-08 Reederey b12x kit (vLLM + FlashInfer b12x, MTP-3, 1M context): 3-rep code bench c1 40.2 / c8 79.4. Same model×engine-family×dual-Spark as this row's 31.2, but a code harness with a pinned image; do not overwrite. Record 40.2 as a code-task note.Measured2026-10-26link
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
ThinkingCap-Qwen3.8-Flash-Nextllama.cpp1× DGX Spark (GB10)30.6code 30.6, prose 30.3; first token after a 32K prompt 51.9 s; aggregate (separate) 4 requests 46.0, 8 requests 44.3; default settings were killed by OOM in 3 of 3 triesMeasured2026-10-26link
Qwen3.8-27Bllama.cpp1× DGX Spark30.02HumanEval+ 50 problems, thinking on, c=1 net decode: plain 10.8 / MTP 22.2 / DSpark 19.7 / DFlash2 30.02 tok/s (tok_s uses DFlash2 code-task net decode; the post has no prose breakdown)Measuredlink
GLM-5.3-FlashUnspecified2× DGX Spark3030 tok/sMeasuredlink
Ornith-1.5-122Bllama.cpp1× DGX Spark29.2code 29.2 (MTP on, 1.22× vs no MTP) / math 30.8; prose is slower with MTP on (repo data); 8K prefill 1,092 (ubatch 4096)Measured2026-10-21link
Ornith-1.5-122BTabbyAPI1× DGX Spark29.2exllamav3 + TabbyAPI (sm_121 build): decode 29.2 (fastest of the three builds); 8K prompt 5.3 s; about 71 GiB memory; GPQA Diamond 70.7 / IFEval 78.5; uses TabbyAPI, not Mia's own recipeMeasured2026-10-21link
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
Ornith-1.5-122Bllama.cpp1× DGX Spark28.2decode 28.2; 8K prompt 8.5 s; about 80 GiB memory; prebuilt CUDA 12.8 image (up to sm_120) gives only 1.8 on GB10Measured2026-10-21link
RED-SNOW-5.3-FLASH (GLM-5.3-Flash refusal-removed version)TabbyAPI1× DGX Spark27.5Part 2 single stream: structured 45.2 / JSON 45.7 / code 35.9 / prose 27.5; 33K research turn 32; 8K→240K decode 50.1→44.4; TTFT 8K 14 s / 65K 79 s / 240K 280 s; prefill 450–640 tok/s; shallow 69 W; part 1 (earlier post by the same author): structured 43.2 / code 40.7 / prose 29.2Measured2026-10-14link
Ornith-1.5-122BvLLM1× DGX Spark26.5decode 26.5; 8K prompt 3.0 s (fastest of the three builds); about 110 GiB memory (vLLM reserves KV); GPQA Diamond 70.2 / IFEval 82.0; same machine Q4_K_M llama.cpp 28.2 (x-099), EXL3 TabbyAPI 29.2 (x-110)Measured2026-10-21link
Clover-35B-A3BSGLang1× DGX Spark26Diffusion LLM. Code 25.7 / prose 26.0; first token 1.4s at 8K, 6.6s at 32K, 40.8s at 128K. SGLang nightly plus author PR 46927 (7-line device allow-list change for GB10). 512 tokens, greedy, thinking off, median of 2. tok_s uses prose 26.0Measured2026-11-05link
Nanbeige4.3-33BvLLM1× DGX Spark22NVFP4 + DFlash2: JSON and code 22–24, prose 11–13; no drafter 11.3 (20 on 2 Sparks); FP8 on 1 Spark 6.3; BF16 on 2 Sparks 6.8. GPQA Diamond / IFEval (thinking off) are within noise for all three precisionsMeasured2026-10-29link
MiniMax-M3.1-FlashvLLM2× DGX Spark21.6Code 21.6 (2.15bpw gives 29.1); IFEval 74.5% vs 60.5%; decode stays flat to 146KMeasured2026-10-28link
GLM-5.3 (744B full, not Flash)Unspecified2× DGX Spark20Prose about 20Measured2026-10-29link
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
GLM-5.3 (744B full, not Flash)TideLM2× DGX Spark (GB10, TP)17.2Median of 3 runs, 1,024 tokens, temp 0.6: prose 17.2 / code 20.8; first token after an 8K prompt 19.5 s. tok_s is proseMeasured2026-10-27link
Muse Spark Open 816B-A42Bllama.cpp2× DGX Spark (GB10, 200GbE RoCE)7.01single-stream decode ~7 tok/s (28-tok prompt 7.01; 650-tok 6.88; 9,146-tok 6.76); load ~10m48s; prefill ~26 tok/s short / ~263 tok/s at 9K; 4 concurrent slots ~14–16 tok/s combinedMeasured2026-10-17link
Nanbeige4.3-33BvLLM1× DGX Spark4about 4Measured2026-10-26link
Qwen3.8-Flash-NextvLLM1× DGX SparkProse 50.7 (1 stream) → 122.9 (4) → 189.6 (8) → 282.6 (16); author's table says 322 at 16 streamsMeasuredlink
Qwen3.8-Flash-NextvLLM2× DGX SparkProse 81.6; structured 138.3; code 133.2; 4 streams structured total 383.1Measuredlink
Qwen3.8-27BSGLangRTX 5090 / 1× DGX Spark206 tok/s / 38 tok/sMeasuredlink
GLM-5.3-FlashTensorFold2× DGX SparkProse 60.4; structured 114.7; 4 streams total 108.8; 8K 54.9 → 874K 44.6; 2026-11-05 MiaAI 1104 recipe is on x-017 (independent reproduction prose 38.7)Measuredlink
MiniMax H3 (Sol accelerated / not accelerated)NVIDIA Sol1× DGX SparkSol: 480p 5 s ≈ 68 s (NVIDIA published 56 s); without Sol >400 s; without Sol 720p 20 steps: 5 s 7 min / 10 s ~20 min / 15 s ~45 minMeasuredlink
MiniMax H3 vs LTX-2.3 ErosUnspecified1× DGX Spark864×480×5s: H3 average 270 s; LTX 154 s (1.76× faster)Measuredlink
Wan2.2ComfyUIDGX Spark compatible system (thermal throttling)736×512 / 16fps: 8 segments ×~5 s (~40 s total) 13.5 minutesMeasuredlink
Ornith-1.5-35B-A3BUnspecified1× DGX Spark24 concurrent sessions up to 440 tok/s (total)Measured2026-08-19link
GLM-5.3-FlashUnspecified3× DGX Spark8 conversations total 214 tok/s, slowest conversation 26–30 tok/sMeasured2026-10-14link
Spark-X2.5/2.6-4BTensorFold1× DGX Spark (GB10)160single stream 160; prefill 5,800 at 2kOfficial2026-10-26link
LFM2.5-24B-A2BTandemLLM1× DGX Spark151EXL3 3.5bpw (default) chat decode 151 tok/s, 11.2 GB, top-1 agreement with BF16 86.7%; another post: plain generation 185, edits up to 350; the same author's other post on 10-15 (UTC+8) claims chat ~300 tok/s (2110434510936723897, conditions not stated, video not checked, not counted in tok_s)Official2026-10-15link
Nemotron-3.5 Lightning 30B-A3BTensorRT Edge-LLM1× DGX Spark 64GB126.9Batch 1, 128-token output, thinking on: HumanEval prompts 126.9 (73.4 without the drafter); SpeedBench 1K 108.5 (73.3); SpeedBench 2K 105.3 (72.8); peak memory 27.9 GBOfficial2026-10-30link
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
Qwen3.8-27BTensorFoldDGX Spark102.9102.9 vs 34.9 tok/s (vLLM+MTP)Relayed2026-09-27link
Qwen3.8-Flash-NextvLLM2× DGX Spark (direct QSFP link, TP2)95NVIDIA cluster playbook command (image vllm-flash-next:8244d1c), fp8 KV: single stream JSON 98 / code 95 / rewrite 75 / explain 74 (median of 5, temperature 0.7); aggregate not measuredOfficial2026-10-25link
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
GLM-5.3-Flashgr82× DGX Spark (GB10) TP2, ConnectX 200G direct link76Recipe-author post: single-stream 76 (“76 if you're just running one”); about 100 with 8 streams is a multi-client total, not the headline. The README body has no speed table and no raw log; claim. Engine gr8, release 901ad084. The post does not name the profile (default MAX_MODEL_LEN 262,144; 1,048,576 with TUNE=context), so context is left null. Recipe repo github.com/kindlingai/glm-5.3-flash-exlr8-2x-spark.Claimed2026-11-10link
Qwen3.6-35B-A3BOllama1× DGX Spark73.5256-token warm start: 1 node 73.5 tok/s; split across 2 nodes via RPC 50.7Official2026-10-15link
Qwen3.8-Flash-NextTensorFold2× DGX Spark67.467.4 / 89.3 tok/s (author says the latter loses about 10% general capability)Relayedlink
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
Qwen3.6-35B-A3BUnspecified1× DGX Spark6565 tok/s; 1,960 PPRelayedlink
Nemotron 3.5 Super VLvLLM1× DGX Spark64JSON 71 / code 64; 29 at 128K context. tok_s uses code. Two changes to the NVIDIA Spark command: memory 0.70, window 256KClaimed2026-11-03link
gpt-oss-120bvLLM2× DGX Spark (TP=2)55.7single request vLLM TP=2 55.7 tok/s (8 concurrent total 202); SGLang TP=2 on same machine 53.3; for comparison: Ollama single node 42Claimed2026-10-15link
Ling-3.0-flashSGLang1× DGX Spark53.89GSM8K, 1,319 questions: decode 37.81 → 53.89 (+42.5%); mean latency 15.21 → 11.49 s; no statistical accuracy changeOfficial2026-10-29link
Qwen3.8-27BTandemLLM1× DGX Spark49.9StairCut speculative decoding before/after 13.7 → 49.9 tok/s (from TandemLLM manual intro post) (whether this number in the manual is measured/computed/estimated not verified)Official2026-10-15link
Beam-1.0-501B-A23BTensorFold2× DGX Spark45.5Code 45.5 (+36% vs vLLM); a short chat after a 46K prompt 2.2 s (was 33 s); ctx 868K (vLLM 655K)Claimed2026-10-29link
Qwen3.8-27BFoundry Local1× DGX Spark44.6Official Microsoft blog chart (DGX Spark): Foundry Local generation speed 4K 44.58 / 16K 43.00 / 32K 41.77 / 65K 40.77 / 131K 34.31 / 260K 30.21; same-setting llama.cpp 27.23 / 24.14 / 26.20 / 20.93 / 15.04 / 12.82. Numbers read from the official chart images. Microsoft's own comparison. tok_s uses the 4K value 44.6Official2026-11-05link
Nemotron-Labs Dory 1.0 4BTensorFold1× DGX Spark (GB10)44single stream 44; aggregate (separate) 8 streams 299Official2026-10-26link
Qwen3.8-27B (AEON Ultimate)vLLM1× DGX Spark (GB10, 121 GiB usable unified memory, DGX OS 7.6.0)43.7vLLM 0.29.0+aeon.sm121a.dspark (image aeon-vllm-ultimate:2026-09-18-v0.29.0-omni), configured context 262,144, FP8 KV, DFlash2 with up to 9 speculative tokens. The author's July script bench_aeon.py: 11 prompts run one by one, greedy, thinking off. Headline is median decode 43.7 (the same in both repeats); min / max 23.7 / 79.4; median TTFT 272 ms. One request with thinking on: 32.3. Quick concurrency test (thinking on) at 1 / 4 / 16 users: aggregate 34.5 / 74.9 / 163.5, not single-stream values. The blog post gives the full config but no raw data, so this is a claim. X post https://x.com/mrru5s3ll/status/2121373870670368890Claimed2026-11-14link
Qwen3.8-Flash-NextTensorFold2× DGX Spark (GB10, 256GB combined, TP2)37.5TensorFold 1.2.5 official release, two-Spark TP2; the author notes this two-box path is experimental. Headline is single-stream prose decode 37.5; code 40.1 is separate. Both are 2-run medians, 512 tokens, temperature 0.7, thinking off. Eight requests at once total 89.0 and is not single-stream. Prefill at 32K is 1,475. Configured window 262,144, 16 GiB prompt cache, 8 slots. No public raw logs, so this is a claim.Claimed2026-11-10link
Gemma 4 12BTensorFold1× DGX Spark (GB10)37single-stream decode 37; with its MTP assistant greedy chat goes 30→67 (2.2×, separate)Official2026-10-24link
Ling-3.1-flash-1019TensorFold2× DGX Spark37Decode about 37; a resent 255K prompt starts output in 0.10 sClaimed2026-10-29link
GLM-5.3-FlashTensorFold1× DGX Spark33.8MTP 1→3 drafts: structured 33.91→40.24, math 32.78→35.67, code 32.67→33.46, prose 27.40→25.79, mean 31.69→33.79. Author headlines 40. tok_s uses the mean 33.8Claimed2026-11-05link
Qwen3.8-Flash-NextSurogate1× DGX Spark31.41 user 31.4 (real text, MTP); aggregate (separate) 16 users 115.7; vLLM needs 123.5 GiB and a Spark has 121.7 GiB, so it does not fit. Another post: Qwen3.8-27B NVFP4 on an RTX 5090, 1-user decode 70.1 (vLLM 71.6)Claimed2026-10-31link
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
Mistral Large 4 (1T-A52B)gr84× DGX Spark, QSFP ring (no switch required)30gr8 @069ec213 + 29 patches in patches/gr8 (commit stated in docs/PATCHES.md). Recipe README table: headline is single-request prose decode 30; code edit 35–40 and new code 31–34 are separate. 20.0 after a 24K prompt. Configured 65,536 (prompt and reply together; one pool for concurrent requests). Two/four-request totals 30 / 39 are not single-stream; with eight requests, four run and four wait, and the total stays 39. tools/bench.py, three fresh starts, recipe-author first-hand table, but with no per-run raw data, so this is a claim. In-window post https://x.com/GrubbyChicken/status/2120597958374912408Claimed2026-11-12link
Gemma 4 31BTensorFold1× DGX Spark (GB10)29.9greedy chat: 12.6 without drafter → 29.9 with z-lab DFlash (2.4×); tok_s is the DFlash valueOfficial2026-10-24link
GLM-5.3-Flash (320B/18B active)Unspecified1× DGX Spark27.6Average 24.6; coding 27.6 tok/sRelayedlink
GLM-5.3-FlashTensorFold1× DGX Spark (GB10, 128GB)26The author's gist linked from the post pins the patches to TensorFold main at 7d97cc656 (after v1.2.6); this is not inferred from the latest commit at post time. Headline is single-stream prose decode ~26; code ~35 is separate; thinking off. Prefill about 650, roughly flat to 128K. Configured as 2 slots of 262,144 each. Two requests total about 34 and is not single-stream. Approximate figures with no raw logs, so this is a claim.Claimed2026-11-10link
MiniMax-M3.1-FlashUnspecified2× DGX Spark23Code 24.7 / prose 23.0; TTFT at 8K 17.7s. tok_s uses prose 23.0Claimed2026-11-05link
Ripple-65B PreviewvLLM1× DGX Spark19.9Code 19.9 / prose 8.2; prefill 1,049 at 2K, 654 at 120K; TTFT 2s at 2K / 34.5s at 31K / 184s at 120K. Early vLLM PR. tok_s uses code 19.9Claimed2026-11-05link
Qwen3.5-397BOllama2× DGX Spark (200 Gb/s direct link, RPC layer split)15.32 nodes 15.3 tok/s (does not fit on a single node), load 188 sOfficial2026-10-15link
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
Ripple-65B PreviewTensorFold1× DGX Spark GB10 128GB10TensorFold 1.2.3 native CUDA ricursive plus PR 2302/2475 stack; not stock release behavior. Single-stream run commit 0c8fd876a. Official target NVFP4 revision 99eea447 and DFlash2 revision 35609b3. Greedy prose 10.0 versus plain 4.4; 256 output tokens, median of three. Main prompt length and full launch CLI unspecified, so classified conservatively as a claim. Greedy code 26.1 is separate; C8 code 49.5/prose 40.3 are aggregate. Cold prefill measured separately at 2048/8192 input; those are not main decode context. Stack CI was not run upstream. On 2026-11-09 the same author/host: single-stream greedy code 24; after draft-cost fix, 8 streams at 15.5 each (was 6.1), 8-stream multi-client code-greedy total 124 (prose-greedy total 62), none replace the single-stream headline. Keep greedy prose 10.0.Claimed2026-11-08link
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
Qwen3.8-27BTensorFold1× DGX Spark (GB10)new default prompt/decode recipe, cold prefill (1.1.3→1.1.4): 2K 834→1,281 (+54%) / 8K 612→1,266 (+107%) / 32K 326→1,165 (+258%) / 64K 203→1,045 (+416%); decode about +2% at 1–8 streams, no absolute decode figure in the notes so tok_s is empty; quality equals the old recipe (27 of 40 exact)Official2026-10-24link
Qwen3.8-Flash-NextTensorFold1× DGX Spark (GB10)prefill only, against 1.1.4: 2k 1,306→1,583 (+21%), 16k 1,359→1,723 (+27%), 32k 1,323→1,687 (+28%), 64k 1,174→1,373 (+17%); served throughput at 1/2/4/8 streams +2/10/20/28%; tok_s is emptyOfficial2026-10-26link
Qwen-Image 2.1SGLang1× DGX Spark (vs 4090)52.6 s / 1024²; SGLang table: Spark 35.4 s, 4090 18.7 sRelayedlink
Qwen3.8-27BTensorFoldRTX 5090 vs 1× DGX SparkNo draft 75 vs 14.4; speculative 215 vs 44 tok/sRelayedlink
GLM-5.3-FlashUnspecified3× DGX SparkProse 77 single stream; 4 streams 146Relayedlink
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
FastH3 V2 / TrimUnspecified1× / 2× DGX Spark1×: 480p 141.4 / 125.8 s, Trim 768p 340.1 s; 2×: 87.2 / 78.3 sOfficiallink
ACE-Step 1.5 + LTX-2.3ComfyUI1× DGX Spark2 min 05 s song generated in one pass; +2:28 lip-sync video end to end 32 minutesRelayedlink

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.