⚡ ローカル推論レシピ

English | 简体中文 | 日本語

RTX 3090 の構成:ローカル推論速度(tok/s)、出典付き

2026-11-14 時点で、このデータセットには RTX 3090 のレシピが 67 件あります。投稿者実測は 33 件です。単一デバイス・シングルストリームのテキスト実測の最速は 384 tok/s(LFM3-1.7B、エンジン llama.cpp、RTX 3090 24GB、実測、2026-10-28、出典)。日付のあるエントリは 57 件(2026-07-22〜2026-11-13)。

エクスプローラーで絞り込む →

モデルエンジンハードウェアtok/s指標出典レベル日付出典
LFM3-1.7Bllama.cppRTX 3090 24GB384Decode about 384, prefill about 21.7K (256K context); 124 with 252K tokens in context; uses 3.9GB VRAM実測2026-10-28リンク
Qwen3.8-27BGeezuz2× RTX 3090 (280W, PCIe, no NVLink)190Reddit post 0-ctx table: Q4 single 190 / 8-request aggregate 622; Q8 143 / 610; Reddit post 10K ctx single Q4 159, 60K 143; repo table (-d 4, fp16 KV, 4 prompts): average 164.0 vs 99.7 for llama.cpp tensor split + MTP 4 (1.64×; layer split gives 74.1 / 72.8); repo batch table (-c 16384): 4-request aggregate 442, 8 requests 625; at 56,017 tokens 105.4 vs 66.6; a comment says the llama.cpp baseline possibly had P2P off実測2026-10-20リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 24GB (400 W, NVLink) + Ryzen 9 3950X + 121GB DDR4173.2One chat (greedy): 173.2 on the build without slots, 143.6 on the best two-slot setup; aggregate (separate) 2 chats 190.4. Local v0.2.9 build with community patches, not an official release; qualification: Strata #5089 (engine 0.1.43, 2× 3090 NVLink) found --peer-device slower than layer split (decode 0.905×, prefill 0.768×); NVLink does nothing with layer split, see comm-251. The same author's 2026-11-08 dual 3090/NVLink test compares patched 0.1.55 and 0.1.58: greedy 1K/8K/128K decode 163.2/177.7/153.1, broadly within about 5% run spread; prefill is 5–13% slower (github.com/Niko1221/Strata/issues/5528). DFlash2 144.3 and normal MTP 167.5 are different arms. Headline 173.2 stays.実測2026-11-01リンク
Qwen3.8-27Bqwen38-megakernel1× RTX 3090 24GB138.5Sampled (thinking on, temperature 1.0, 8 runs) 138.5 (best llama.cpp on the same card 72.8); greedy code 232.9 / short story 90.8; 109.2 at 100K context (llama.cpp 54.4); 66.6 at 261K deep; 7.5% faster on short prompts with drafting off; greedy output is the same at any draft count; aggregate (separate) 4 agents 382 (llama.cpp 137). On 2026-11-07 the same author posted a new table (v2026.11.05, default 7 drafts, q8_0 KV): short greedy 166.5, sampled thinking 146.9, 116.4 at a filled 101.6K; 4-agent aggregate 424 (separate). preset F is 160.7 / 147.4 / 123.2 and changes the output, so do not mix it with q8_0. Versus llama.cpp b12024 against its own no-draft run, greedy output differs on 113/282 items (score 231→229, 6 correctness flips); the author's engine has 0 draft/no-draft difference, which does not mean it is bit-identical to llama.cpp. tok_s stays the earlier sampled-thinking 138.5実測2026-11-02リンク
Qwen3.8-Flash-NextStrata1× RTX 3090 Ti 24GB + Threadripper 3990X + 128GB DDR4138Decode about 138–140実測2026-11-01リンク
Qwen3.8-27BvLLM4× RTX 3090 24GB (gpu-unlock enables fp32 accumulate)131.9TP4: 1 request 131.9 (128.5 before unlock); 8-request aggregate 571.7 (448 before); about 60K prompt 38.6実測2026-10-23リンク
Qwen3.8-Flash-NextStrata4× RTX 3090 24GB (320W/card, no NVLink) + Threadripper 3990X + 251.5GiB RAM127.4Strata v0.2.12 (146df3c4), engine 0.1.58 CUDA 13 source build, S=setup defaults: 128K, batch 3, int8 KV, MTP. Standard-protocol short single-stream decode median 127.4 (run 2 within 2%); 4K/16K decode 128.3/127.9, prefill 1921/3219. Unmerged PT short decode 154.3 is not the headline. Multi-client c4 totals 198/202/196 and 16-user short total 539.0 (C arm) are not single-stream. 21/21 requests, sanity 12/12; six results/outputs/engine.log artifacts public.実測2026-11-09リンク
Qwen3.8-27Belpis (ExLlamaV3 patch)RTX 3090 24GB (350W)125.7elpis (patched ExLlamaV3 server): decode 125.7 over 1K–262K, prefill 1,220.7 over 8K–262K; GSM8K 40 questions 205.3. int8 attention changes the outputs (lossy); the lossless version gildrb/elpis-lossless gives decode 82.7実測2026-11-03リンク
Qwen3.8-Flash-NextStrataRTX 3090 Ti 24GB + i9-12900KF + 64GB RAM (Windows 11)124Strata 0.1.49.2 @7dd5cf18, engine 0.1.47, strata_bench --long: short chat decode 124.0 (99.2–148.1); agent follow-up 138.1; prefill 4K 2,251 / 16K 2,833 / 32K 2,806実測2026-10-25リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 24GB (NVLink NV4) + Threadripper PRO 3945WX + 128GB 8-channel DDR4 (ThinkStation P620)121.8Engine 0.1.43, layer split against --peer-device: 8K thinking off decode 121.8 against 112.0; 32K 115.8 against 110.0; 100K 107.4 against 98.2; peer mode gives 0.905× decode and 0.768× prefill (about 0.65× at 100K). NVLink traffic is 0 with layer split. This contradicts the older issue #204実測2026-11-03リンク
Qwen3.8-Flash-NextStratasingle RTX 3090 24GB + 64GB RAM (RAM see same author's earlier post)120.8120.8 tok/s live; 2,000 PP; earlier post by the same author on the same day: decode 72 after a new Strata release (128K, single 3090, x.com/needmorevram/status/2104898095318544540)実測リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 (48GB) + 64GB DDR5-6000120Sustained decode 120+ tok/s, peak 153.4; cold prefill 2,844 tok/s (191K), peak 3,113実測2026-10-11リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 + Ryzen 7 9700X + 128GB DDR5117.5Strata 0.1.48.1 @7ddc17a3 (engine 0.1.48), --spec 4 --mtp rt: short chat 117.5 (85.9–134.7, median of 9); prefill 4K 1,292 / 16K 2,416; agent follow-up turns 120.8; qualification: from Strata 0.2.10, Auto defaults to DFlash2; dual-card decode can lag MTP or a single card (Strata #5206: DFlash2≈MTP on 2× 4060 Ti; #5172: dual RX 9070 10–50% slower than one card)実測2026-10-24リンク
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 62実測2026-10-31リンク
Qwen3.8-Flash-Next (abliterated)StrataRTX 3090 24GB (300W) + 41GiB RAM (experts in system RAM)911 card 4K 91–102 / 32K 99–101 / 131K 100–103; prefill 0.9–1.6K; same rig IQ3_S GGUF 74–84; quality data (KL 0.040) from Mirai, not checked by author; 2026-11-02, same tester: 88 at both 32K and 130K, 32K prefill 1,932; 2 cards 123 (x.com/superalesha/status/2117135114081837264)実測2026-10-20リンク
Qwen3.8-Flash-NextStrata1× RTX 3090 + 64GB DDR490.7Same machine, engine 0.1.45 → 0.1.56: decode 69.6 → 90.7 (+30%); agent decode 85.2 → 121.5; first token 3.56 → 2.61 s; prefill no change; same-class Strata 0.2.10: web-admin-master short-prompt decode 97–100 (up from 69–79 on 0.1.40.1), but the host hard-reset 3 times while reading a ~64K prompt; rolled back to 0.1.40.1 and was fine (Strata #5241)実測2026-11-03リンク
Qwen3.8-Flash-NextStrataRTX 3090 24GB + i9-10900K + 64GB RAM (Windows 11)86.9Strata 0.1.48.1, engine 0.1.48, k8v4 KV + --kv-grow, 512K context: short chat decode 86.9; agent follow-up 94.4; prefill 4K 1,102 / 16K 1,675. Engine 0.1.50 rerun: 86.9→99.0 (different day, ranges overlap). Of 7 A/B arms only STRATA_PF_FUSED=1 is beyond noise (+3.6% 16K prefill)実測2026-10-25リンク
Qwen3.8-Flash-Next未指定RTX 3090 24GB (EVGA FTW3) + Xeon E5-1680 v2 (2013) + 128GB quad-channel DDR3-1866 ECC70decode 70 / prefill 1,900 (author says the 3090 does most of the work)実測2026-10-22リンク
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)実測2026-11-03リンク
Nanbeige4.3-33BvLLM2× RTX 3090 + NVLink66single stream 66; the recipe has a C1–C6 concurrency bench実測2026-10-26リンク
Qwen3.8-Flash-NextStrataRTX 3090 24GB + 128GB RAM (Windows)65.5Short-chat decode median 65.5 tok/s (49–81), first token ~0.40 s; agent turn at ~14.5K context 74.8; prefill 4K 1,473 / 16K 2,127 tok/s実測2026-10-14リンク
Nanbeige4.3-33B BasevLLM1× RTX 3090 24GB39.6single stream 39.6 (Qwen3.8-27B on the same machine 27.2); the test uses the Base model実測2026-10-26リンク
DeepSeek-V4.1-Flash-1017Strata1× RTX 3090 Ti 24GB + Threadripper 3990X + 128GB DDR435Decode about 35 (MTP/DSpark on)実測2026-11-01リンク
DeepSeek-V4.1-Flash-1017未指定2× RTX 3090 + 256GB 8-channel DDR4351M context: decode about 35, prefill about 900. The post does not name the engine実測2026-11-03リンク
Qwen3.8-Flash-Next未指定M4 Mac mini 16GB + RTX 3090 (NVMe streaming)31EXL3 3-bit, more than half the model streamed from NVMe: 31 tok/s; post from the day before on the same machine (2026-10-14, quant not stated, x.com/0xSero/status/2110103539813077168): decode 23–27, prefill 523, 200K KV, startup 118 s実測2026-10-15リンク
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%実測2026-10-18リンク
Qwen3.8-Flash-NextStrataRTX 3090 + 32GB RAM (single card) / +2070S (32GB combined)2020–21 tok/s / 34.2 tok/s実測リンク
GLM-5.3 744BProject Maya4× RTX 3090 24GB (ASUS TUF OC, 250W cap) + EPYC 7702 + 512GB DDR4 ECC11.1Maya v1.1.57 (commit 0065548). --bench decode mean 11.1 (9.6/12.2/11.5). Prefill 145 at 2K / 265 at 8K. Context 262144. Full maya-report is in the issue. Consumer 3090 cards; no rack evidence.実測2026-11-10リンク
Kimi K3未指定2× RTX 3090 + 1TB DDR4 (8×128GB)8.5100K context: average 8.5–9; prefill 35. The post does not name the engine実測2026-11-03リンク
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 use実測2026-10-21リンク
Z-Image Turbo (6B, 8 steps)ComfyUIRTX 3090 24GB2048px about 8 s / image; another user 12 s after warm-up実測リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 + 80GB DDR4 (used HP Z840)100+ tok/s; 900–1,200 PP; qualification: from Strata 0.2.10, Auto defaults to DFlash2; dual-card decode can lag MTP or a single card (Strata #5206: DFlash2≈MTP on 2× 4060 Ti; #5172: dual RX 9070 10–50% slower than one card)実測リンク
MiniMax H3ComfyUIRTX 309011 s video 13 minutes実測リンク
Qwen3.8-27BtqLLMRTX 3090 24GB273Short prompt (LocalMaxxing reasoning) 273 tok/s, median of 5, temperature 0, accepted by LocalMaxxing; code prompt by context: 32K 203 / 128K 157 / 192K 144 / 256K 107 (tok_s = short prompt). On 2026-11-09 the author tried DSpine: TQ4_L code/math 209 (same-condition DFlash2 227); TQ3_4S 133 vs DFlash2 162; 5060 Ti 16GB TQ_XS 85 vs 130. DSpine accepts more tokens but draft steps cost more. Keep this row's TQ3_4S·DFlash2 short-prompt 273. On 2026-11-09 the author added prose/long-text results (same 3090, TQ3_4S, greedy, thinking off): DFlash2 243 to 131, DSpine 194 to 109 (draft accept 36% on prose vs 78% on code). Headline 273 is the short-prompt reasoning/code figure; keep it.公式2026-10-15リンク
Qwen3.8-Flash-NextStrataRTX 5090 32GB (PCIe 5.0 x8) single-card engine + Ryzen 9 9950X + 96GB DDR5-6000 (two more RTX 3090 cards in the same box run one engine each; three engines share an expert arena of about 50GB in RAM)160Strata self-built v0.2.12 + PR #3341 + #5203 (engine 0.1.58.2), control 0.1.58 + #3565. Three engines share one expert arena in RAM (--shared-expert-arena, about 50GB), --ple-io direct. This row is the 5090 engine: 512K (YaRN x2, KV q4_0), --batch 2 --batch-mtp, 7 cores. Greedy short prompts: decode alone 160-205; the headline is the low end, 160. With all three engines busy: 139-186. Live traffic, 15.5K requests: decode median 164 (p90 192). Box-wide 4 streams give 371-392 aggregate, not a single-stream value. No raw logs, so this is a claim. The 3090 engines are comm-394.主張2026-11-13リンク
Qwen3.6-35B-A3B-Escha-W2SGLangRTX 3090 24GB154154 tok/s decode主張2026-07-22リンク
Qwen3.8-Flash-NextStrata4× RTX 3090 24GB (275W cap, no NVLink, PCIe 3.0) + Threadripper 2920X + 128GB (GPU1/3 x16, GPU0/2 x8)139Strata 0.2.13 (engine 0.1.59), layer split, DFlash2 drafter, int8 KV, 32K KV cells per layer in VRAM, control vector on layers 4-44. Thinking off, temperature 0, one request at a time, each config started fresh and run twice. 4 cards (0+1+2+3) short decode 139-144; the headline is the low end, 139. After 50K: 85-96. Cold read of 2K/16K/50K: 2.0/6.4/12 s. Expert hit 99%. The author notes that accepted drafts change tok/s. Summary tables only, no per-run raw data, so this is a claim. 3-card and 2-card rows from the same post: comm-386, comm-387.主張2026-11-13リンク
Qwen3.8-27BDFlashRTX 3090 24GB13882 native → 114 MTP → 138 DFlash2 tok/s伝聞リンク
Recursive Flint 40B-A2.5Bllama.cppRTX 3090 24GB130.8chat-short decode 130.8 (8K) / 128.9 (16K); 16K: chat-long 123.0 / prefill 2,740, agent-trace 124.4 / prefill 6,081; same card and build Qwen3.8-27B Q4_K_M 38.83 (3.37×). UD-Q4_K_XL (25.05GB) does not fit in 24GB: about 17 with part on the CPU; 1 run per configuration主張2026-10-22リンク
Qwen3.8-Flash-NextStrata3× RTX 3090 24GB (275W cap, no NVLink, PCIe 3.0) + Threadripper 2920X + 128GB (GPU1+2+3: x16/x8/x16)112Same machine and settings as comm-385, with 3 cards (1+2+3). UD-Q4_K_XL short decode 112-137; the headline is the low end, 112. After 50K: 100-105. Expert hit 95%. Other 3-card orders read 119-134; the author says the difference is inside noise. UD-Q5_K_XL on the same 3 cards: 91-93 (PPL 2.4781 against 2.4919 for Q4), note only. 200K chat: a cold 189,809-token read takes 91 s, then a short question decodes at 78. Summary tables only, no per-run raw data, so this is a claim.主張2026-11-13リンク
Qwen3.8-Flash-NextStrataRTX 3090 24GB (PCIe 4.0 x8) single-card engine + Ryzen 9 9950X + 96GB DDR5-6000 (same box: an RTX 5090 and a second RTX 3090 on OCuLink PCIe 4.0 x4 run one engine each and share the RAM expert arena)110Same machine and build as comm-393. This row is the 3090 engine on PCIe 4.0 x8: 256K, 1 slot, 5 cores. Greedy short prompts: decode alone 110-120; the headline is the low end, 110. With all three engines busy: 98-111. The 3090 engine on OCuLink x4 reads 113-127 alone and 90-125 with all busy; widened to 9 cores (8 workers) it reads 134-136 alone, but with both 3090 engines widened both fall to 27-47. Notes only. No raw logs, so this is a claim.主張2026-11-13リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 24GB (one PCIe 4.0 x8 main card; one PCIe 3.0 x1 riser card as helper) + Ryzen 5 8600G + 61GiB (Arch Linux)99Strata engine 0.1.59.1, local build (v0.2.13.1-88-gdbb2d168). Layer split 15, STRATA_STAGE_TRIM=1, --pipeline-windows 2 + STRATA_MTP_STAGE0=1, --spec 4, --kv int8, --max-context 262144, Q8_0 PLE with --ple-io direct. Cold-probe decode 99-110; the headline is the low end, 99. Real chats on this config: median 100.8 over 57 requests (77-123). Cold read at 20.7K: about 2,122. The x1 riser card is the helper with layers 0-14. The author says the raw zip is attached only on request; it is not public in the issue, so this is a claim.主張2026-11-12リンク
Qwen3.8-Flash-NextStrata2× RTX 3090 24GB (275W cap, no NVLink, PCIe 3.0) + Threadripper 2920X + 128GB (GPU1+3, both x16)96Same machine and settings as comm-385, with 2 cards (1+3, both x16). UD-Q4_K_XL short decode 96. After 50K: 83-90. Expert hit 88%. The x8 pair (0+2) decodes at 90-95; the author says PCIe width changes prompt reading, not decode. UD-Q5_K_XL on the same 2 cards: 68-69, note only. Summary tables only, no per-run raw data, so this is a claim.主張2026-11-13リンク
Qwen3.8-Flash-Next (abliterated)Strata1× RTX 3090 24GB91Strata 0.2.10 + DFlash2, 16K context: code 110 / English 91, prefill 1.8K. 2×3090: code 151 / English 126, prefill 2.6K; at 128K dual-card code still 147. Quant file is Unsloth's 11-04 Dynamic V3.1 re-upload (d9620119); Strata #5296 notes that file is 40–55% slower at prefill on gfx1151 — this row is CUDA 3090. tok_s uses English 91主張2026-11-05リンク
Nemotron 3.5 Bolt 74B-A9Bllama.cpp1× RTX 3090 24GB87.8Out-of-window makeup (measured 2026-10-15, cited in-window by @llmspeed). Signed llm-speed run r_-sifd5otbpy, suite-v1, llama.cpp b11669. chat-short decode 87.75, recorded as 87.8 (prefill 1,183.4, TTFT 462 ms). The blog states the measurement is at 4K context. No 3090 sibling in the baseline (x-271/x-272 are 4090/5090). Page https://llm-speed.com/blog/nemotron-3-5-bolt-74b-24gb-measured The llm-speed page shows only a summary and a signature, with no raw records and no full server settings, so this is a claim.主張2026-10-15リンク
Qwen3.8-Flash-NextStrata1× RTX 3090 24GB (280W cap) + Xeon Gold 6248 + 192GB DDR4 ECC65.95Strata v0.1.58 (engine telemetry; exact Git commit not recorded). Custom long-context reasoning suite, not standard strata bench. Headline is the 12-request decode telemetry median 65.95 (59.5–83.5). Context 524,288, extended from 262,144 with YaRN 2x. The write-up says Zagitta/gpu-unlock was reported enabled, but the revision and boot evidence were not archived and there is no A/B, so no gpu-unlock tag. No live JSON, so this is a claim. No rack evidence.主張2026-11-11リンク
Qwen3.8-Flash-Next未指定RTX 3090 24GB + 64GB DDR4/5 + 100GB NVMe64c1 decode 64 tok/s; prefill 2,000 / 2,800 tok/s主張2026-09-29リンク
Qwen3.8-27BvLLM1× RTX 3090 24GB (dedicated; display on a 3090 Ti)48.7vLLM v0.31.0, --gpu-memory-utilization 0.94. Author's post: this AWQ and the same-recipe abliterated build both reached about 40K KV cache and 48.7 tok/s. The stock AMD Quark build ran out of memory during weight conversion on 24GB and did not load. Honcho JSON 200/200. The ~40K figure is measured KV use, not a stated configured limit; context is left null. No raw log; claim.主張2026-11-10リンク
Qwen3.5-27BOllama1× RTX 3090 24GB42Ollama 0.40.15, default settings, one request: 42. Four requests at once: default aggregate 39.8; OLLAMA_NUM_PARALLEL=4 aggregate 33.1 (32K context per request, 16% of the model moved to the CPU); these are not single-stream values. In the same post llama.cpp also gives 42 for one request and 100.2 aggregate for four, but the llama.cpp version is not stated, so it has no row. The context limit for one request is not stated, left null. No raw logs, so this is a claim. Full write-up https://homebrewserver.club/ollama-vs-llama-cpp.html主張2026-11-13リンク
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.主張2026-11-05リンク
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 41公式2026-10-25リンク
Qwen3.8-27B-S-miraillama.cpp (mirai-s fork)RTX 3090 (VRAM limited to 12GB, PCIe 3.0)39.739.7 tok/s decode; 1008 tok/s prefill主張2026-09-26リンク
Qwen3.8-27Bllama.cpp1× RTX 3090 24GB38.61Out-of-window makeup (measured 2026-10-22; cited in-window by @llmspeed). llm-speed signed run r_mmtzel_i367, suite-v1, llama.cpp b11803. chat-short decode 38.61 (prefill 482.3, TTFT p50 25.8 ms); chat-long 38.19. Context 4K. No baseline row with this llama.cpp × Q4_K_M × 3090 × Qwen3.8-27B combo. The llm-speed page shows only a summary and a signature, with no raw records and no full server settings, so this is a claim.主張2026-10-22リンク
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 → 1,840主張2026-11-03リンク
Qwen3.8-27B + Tencent DSpine-b16llama.cpp (Ajay9o9 dspine fork)1× RTX 3090 24GB35.6Ajay9o9/llama.cpp@dspine (unreleased). llama-benchy novel text: 8K-depth DSpine decode 35.6 (vs DFlash2 26.6 / no-spec 28.1); 32K/64K 31.1/35.5. Code/math ~125 and mixed chat 90 are task-skewed and not the headline. About 64K with q8_0 KV; about 98K at n-max 7. No raw log and no full CLI; claim.主張2026-11-09リンク
Qwen3.8-Flash-Next未指定RTX 3090 24GB + 64GB DDR431.7average about 31.7 (simple questions, reasoning OFF / MAX); RAM 56–57 / 64GB, VRAM 19.8 / 24GB; context 16K / 32K主張2026-10-25リンク
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.主張2026-11-13リンク
Qwen3.8-Flash-Nextllama.cppRTX 3090 + 64GB RAM2828 tok/s (sustained over 1.59M tokens) / 38.6 tok/s伝聞リンク
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 26.0主張2026-11-05リンク
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)主張2026-11-04リンク
MiniMax-M3.1-FlashProject Maya4× RTX 3090 24GB + EPYC (2 of 8 memory channels filled)23Decode 23–31 (best llama.cpp on the same machine 6.7); prefill 16K 1,120 / 28K 1,525 (llama.cpp 219); cold experts through the CPU at about 19 GB/s主張2026-10-28リンク
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 mean主張2026-10-27リンク
Kimi K3.1 ProProject Maya3× RTX 3090 + EPYC 9124 + 128GB + 1 NVMe1Project Maya open PR #962 (maya-k31, based on v1.1.60, head fa283b528; a separate engine, off by default). 3×3090 plus NVMe-streamed experts, IQ1_S decode 1.0 as the headline. Top-1 agreement with BF16 is 92.1%. --n-ctx was tested to 32K, which is a test ceiling, not a stated configured limit, so context is left null. No rack evidence; recorded as a workstation. Figures are in the PR body, no raw log; claim.主張2026-11-12リンク
Kimi K3.1 ProProject Maya3× RTX 3090 + EPYC 9124 + 128GB + 1 NVMe0.88Same PR, machine and engine as comm-364, mixed IQ1_S + IQ2_XXS quant, decode 0.88, so this is a new row. Test ceiling 32K; configured context limit is unstated, left null. No raw log; claim.主張2026-11-12リンク
Qwen-Image 2.1未指定RTX 309082–96 s / 2K (public receipt)伝聞リンク
Qwen3.6-35B-A3BOllamaRTX 3090139.6 @32K; 112.4 @64K (0.71GiB spilled to CPU)伝聞リンク
ACE-Step 1.5 (music)未指定RTX 3090Whole song <10 s (promo post)伝聞リンク

数値は各出典のものです。本プロジェクトでは再計測していません。行ごとにモデルと設定が異なります。出典レベル:公式 = プロジェクトまたはベンダー、実測 = 投稿者が実行(エンジンまたはレシピの作者は生データと再現手順も公開する)、伝聞 = 二次情報、主張 = 予測または再現手順なし。