RTX PRO 6000 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 PRO 6000 のレシピが 21 件あります。投稿者実測は 13 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 300 tok/s(Qwen3.8-27B、エンジン SGLang、RTX PRO 6000、実測、2026-08-20、出典)。日付のあるエントリは 21 件(2026-08-20〜2026-11-13)。
| 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 |
|---|---|---|---|---|---|---|---|
| Qwen3.8-27B | SGLang | RTX PRO 6000 | 300 | about 300 tok/s (post); mia-ai.net lists 240+ at 8 requests | 実測 | 2026-08-20 | リンク |
| Qwen3.8-Flash-Next | vLLM | 1× RTX PRO 6000 Blackwell 96GB (450W cap) | 287 | Coding tasks (about 250 tokens in, 1,536 out, temperature 0, thinking off): single stream 287 (wall clock, prefill included); prefill about 14.5K at 4K, about 13.5K at 128K; aggregate (separate) 8 concurrent 965 | 実測 | 2026-10-30 | リンク |
| Qwen3.8-Flash-Next | TensorFold | 1× RTX PRO 6000 Blackwell 96GB | 271.4 | Prose decode at 400 W (temperature 0.7, thinking on): 1 request 271.4; aggregate (separate) 2 requests 395.7, 3 requests 503.7; prefill at 600 W 4K 9,218.8 / 32K 8,519.9 / 128K 7,886.9. The X post says 600 W, but the recipe page says decode was measured at 400 W. This entry uses the recipe page | 実測 | 2026-11-02 | リンク |
| Recursive Flint 40B-A2.5B | llama.cpp | 1× RTX PRO 6000 Blackwell 96GB | 244 | 1 request 244; 135 per request with 3 at once; prefill 5,230; one run | 実測 | 2026-10-23 | リンク |
| Swift-1.5-Qwen3.8-Flash-Next | ninfer | RTX PRO 6000 Blackwell 96GB @ 600W + 96GB DDR4 + Ryzen 5 5600X | 238 | decode mean ~238 tok/s (Ninfer6000 blog sweep mean; MTP 3 draft) | 実測 | 2026-10-15 | リンク |
| Qwen3.8-Flash-Next | Strata | 1× RTX PRO 6000 Blackwell WS 96GB (350W cap of 600W) + Windows 10 | 228 | Strata v0.2.12 plus local logging branch 3feea4f4 (engine unchanged), engine 0.1.58 official Windows CUDA 13. Short-chat median decode 228 (209–327); agent 266; prefill 4K/16K/32K 4446/5054/5126. Two slots, int8 KV; needle 6/6 at 32K/128K. results.json and needles.json public. Hard 350W power cap. Separate from same-card UD-Q2_K_XL row x-293. | 実測 | 2026-11-08 | リンク |
| Qwen3.8-Flash-Next | SGLang (Penny Royal fork) | 1× RTX PRO 6000 Blackwell 96GB (300 W limit) | 200 | Penny Royal 3.1.4, short prompts, 400-token generations: 1 stream about 200 (stock SGLang about 94 on the same card); aggregate (separate) 4 streams about 500, 8 streams about 554 (74 per stream); prefill about 12.5K; KV pool 942K tokens | 実測 | 2026-11-02 | リンク |
| Qwen3.8-Flash-Next | vLLM | RTX PRO 6000 Blackwell 96GB @ 600W + 96GB DDR4 + Ryzen 5 5600X | 150 | decode mean ~150 tok/s (blog sweep mean; MTP 2) | 実測 | 2026-10-15 | リンク |
| Qwen3.8-Flash-Next | llama.cpp | RTX PRO 6000 Blackwell 96GB @ 600W + 96GB DDR4 + Ryzen 5 5600X | 108 | decode mean ~108 tok/s (blog sweep mean; MTP 3) | 実測 | 2026-10-15 | リンク |
| Nanbeige4.3-33B | 未指定 | 1× RTX PRO 6000 Max-Q (325 W) | 105.3 | Batch 1: NVFP4 + n-gram code 94.1, reasoning 92.6; with DFlash2 code 105.3; stock FP8 42.9. tok_s is the DFlash2 code value | 実測 | 2026-11-02 | リンク |
| DeepSeek-V4.1-Flash-1017 | Strata | 1× RTX PRO 6000 Blackwell Workstation 96GB + Ryzen 7 9800X3D + 61GB RAM (Linux) | 86.1 | short decode 86.1 (58.5–130.9); 32K prefill 2,185.6; Linux. For Windows data on the same card, see the other entry | 実測 | 2026-10-26 | リンク |
| Ripple-65B Preview | SGLang | 1× RTX PRO 6000 Blackwell 96GB | 49.3 | Target SAGE-REXL3 5.41bpw, with a Qwen3-8B EXL3 4.55bpw DFlash2 draft. SGLang 0.5.19 + sglang-exl3 plugin 5d1c9ef. Greedy: peak 64.8 on a 403-token short prompt; 44.9 at 1,564 tokens; 49.3 at 3,152 tokens. Sampled peak 53.1. Cold prefill about 1,796. tok_s uses the longer-generation 49.3, not the 403-token peak 64.8. The 22.3GB / 2.57bpw pack has no matching speed and is not reused. 99.34% is next-token top-1 agreement over 55,535 assistant-turn tokens, not task accuracy. Same model and card as x-259 but a different quant/engine/method; x-259's 135.2 is unchanged. Recipe: github.com/vcruz305/Ripple-65B-Preview-EXL3-SGLang-recipe | 実測 | 2026-11-07 | リンク |
| DeepSeek-V4.1-Flash-1017 | Strata | 1× 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 empty | 実測 | 2026-10-26 | リンク | |
| Qwen3.8-Flash-Next | Strata | 1× RTX PRO 6000 Blackwell 96GB | 300.3 | Strata 0.2.12 / engine 0.1.58 / CUDA 13. Short-input median decode 300.3; 7K/29K 311.9/271.2; prefill 935.8/7088.6/7561.1. Three runs × 512 output tokens per depth, zero cached input. Configured context 256K, int8 KV; deepest measured input 29K. No raw results file and no full CLI; host CPU/RAM omitted; classified as a claim under the raw-data rule. Separate from 350W-capped NVFP4 row comm-300. | 主張 | 2026-11-09 | リンク |
| Qwen3.8-Flash-Next | ExLlama | 1× RTX PRO 6000 Blackwell Max-Q 96GB | 299.4 | V2 recipe (repo commit a659d52056): exllamav3 v1.6.4 + engine/patches (V2 patches) + TabbyAPI @564f8fb, 6-bit KV, 3M-token shared pool, up to 655,360 tokens per request, 8192-token prefill chunks, DFlash2 drafter. Code prompt, 512 output tokens per request, greedy, A/A/B/B alternating server boots, 3 runs per boot, medians. Headline is 1 user, 299.4 (measured with a 3,121,152-token pool and 2048-token prefill chunks). 2 / 4 / 8 users aggregate 398.0 / 560.9 / 676.7, not single-stream values. The shipped config gave 305.3 at 1 user in a single run. Decode of 2048 forced tokens from a short prompt: prose 125.8, code 287.3, chat 156.1, math 313.8. Prompt read 32K / 128K: 5,690 / 5,647. Old headline: v1 (stock exllamav3 1.6.4, 8-bit KV) code prompt C=1 233.9 (README of 2026-10-24, median of 3, same repo URL); v1 in the V2 table measures 231.1. The repo gives methods and median tables but no per-run raw data, so this is a claim. X post https://x.com/GPTWare/status/2121265574529339808 | 主張 | 2026-11-13 | リンク |
| GLM-5.3-Flash | 未指定 | 2× RTX PRO 6000 Blackwell Max-Q 96GB (300W cap) | 232 | Reproduction of the MiaAI_lab recipe, prose, temperature 0.7, thinking on, 400 tokens, average of 2 runs: 300W C1 232; 275W (stock) C1 207. Aggregate (separate) at 300W C2 329 / C3 384 / C4 442; 275W 303 / 360 / 406. TTFT 47 ms at C1 | 主張 | 2026-11-03 | リンク |
| Ripple-65B Preview | 未指定 | 1× RTX PRO 6000 Blackwell 96GB | 135.2 | LocalMaxxing verified page (batch 1, temp 0, median of 5): code 135.2, reasoning 97.9, 24.0 without the drafter. tok_s uses code 135.2 | 主張 | 2026-11-05 | リンク |
| DeepSeek-V4.1-Flash | Tandem-vLLM | 1× RTX PRO 6000 96GB + EPYC 9534 | 124 | Tandem-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/ | 主張 | 2026-11-12 | リンク |
| Recursive Flint 40B-A2.5B | vLLM | RTX PRO 6000 Blackwell 96GB | 110 | single stream about 110–180; 64 concurrent requests peak about 2,500 total; author ran 45 LiveCodeBench v6 problems (82.2) and 50 SWE-bench Verified tasks (80) (subsets, not full evals) | 主張 | 2026-10-22 | リンク |
| Qwen3.8-27B | SGLang | 1× RTX PRO 6000 | 110 | SGLang 0.5.19, MiaAI-Lab RTX PRO 6000 recipe (the recipe repo is named for DSpark). Same machine, quant and engine; the three drafter configs are one row. Headline is the recipe default, DSpark, short-prompt prose decode (greedy, median of 3, 512 tokens) 110; same-method DFlash2 108 and no draft 58 are in this row. Short-prompt code: DFlash2 171 / DSpark 172; prose at about 29K: DFlash2 59 / DSpark 57 / no draft 40. Configured context is unstated, left null. Method is described, no raw log; claim. | 主張 | 2026-11-11 | リンク |
| GLM-5.3 (753B full, not Flash) | 未指定 | 1× RTX PRO 6000 Blackwell 96GB (about 250GB offloaded to system RAM) | 45 | One card 45 / prefill 1,200; TP2 on two cards 62 / prefill 1,600. No official drafter. tok_s uses one-card 45 | 主張 | 2026-11-05 | 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。