⚡ ローカル推論レシピ

English | 简体中文 | 日本語

RTX 5060 Ti 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 5060 Ti のレシピが 27 件あります。投稿者実測は 20 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 162 tok/s(LFM2.5-24B-A2B、エンジン llama.cpp、RTX 5060 Ti 16GB、実測、2026-10-15、出典)。日付のあるエントリは 26 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
LFM2.5-24B-A2Bllama.cppRTX 5060 Ti 16GB162162.0 tok/s; math problem 3.99 s; GSM8K 97%, MMLU 82%; same GGUF, same llama.cpp build; writing speed median of 5 runs実測2026-10-15リンク
Flint 40Bllama.cppRTX 5060 Ti 16GB115.8llama.cpp b11972, median of 2 runs of 512 tokens: prose 115.8 / code 116.0; 90.7 at 32K; TTFT 15.6 s at 32K. RTX 5090 NVFP4 for comparison: 224.9実測2026-11-03リンク
Qwen3.8-Flash-NextStrata1× RTX 5060 Ti 16GB + 61GB RAM95Decode about 95–106 across about 100K context; warm cache re-read about 0.2 s; Basalt gives +35% prefill, Strata gives +15% decode; VRAM 15.8/16 GB; qualification: another 5060 Ti 16GB tester was 4–7% slower after the update to Strata 0.2.10 (engine 0.1.54 → 0.1.56, UD-IQ4_XS): chat 32.1 → 30.3, agent generation 39.5 → 38.1 (x.com/yutacar/status/2117447921100833121)実測2026-11-01リンク
Qwen3.8-Flash-NextStrataRTX 5060 Ti 16GB (PCIe 5.0 x8) + RTX 3060 12GB (chipset PCIe 4.0 x4) layer split + Core i5-14400F + 62.6GB DDR5 (Linux)87.9Strata 0.1.59.1 (0.2.13.1) plus three unmerged conversation-cache commits (the author says they do not touch the prompt or decode path). Configured 262,144, int8 KV, layer boundary 17. Headline is reasoning decode with MTP (--spec 4), median 87.9 (75.8–88.1); short decode on the same setting is 75.2. DFlash2 reasoning 77.6 is separate. The author says the 3060 used gpu-unlock 1.0.0, so the row gets the gpu-unlock tag per repo convention. Per-run jsonl is in the PR.実測2026-11-12リンク
Qwen3.8-Flash-NextStrata2× RTX 5060 Ti 16GB (layer split, PCIe 3.0 x8 / x4) + 6-core Zen 2 CPU79.6Short-chat decode 79.6; agent turns 92.4; prefill 4K 696.5 / 16K 1,133.3 / 32K 1,371.8; conversation cache hit 69.3%; 3 runs each; 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2026-10-29リンク
Qwen3.8-Flash-Next AbliteratedStrataV100 32GB + RTX 5060 Ti 16GB + 32GB DDR57878 tok/s @32K; 56 tok/s @262K実測リンク
Qwen3.8-27BExLlama2× RTX 5060 Ti 16GB73.1code 73.1, prose 68.4; 32K prompt read 996実測2026-10-26リンク
Qwen3.8-Flash-NextStrataRTX 5070 12GB + RTX 5060 Ti 16GB (two different Blackwell cards, layer split) + Ryzen 9 9950X3D2 + 125GB RAM (Windows 11)67.3After tuning (STRATA_ARENA_PIN_GIB=40, STRATA_STAGE_TRIM=1, layer split K=28, Cyrillic draft vocabulary): decode 67.0 on a Russian prompt and 67.3 on an English prompt; prefill 2,161. Before tuning: decode 46.2, prefill 1,567.7. STAGE_TRIM alone gave 53.9. 6 runs per point; the author used own prompts, not strata bench. tok_s is the English value; on 2026-11-02 the same tester swept engines 0.1.47–0.1.55 (NVFP4 GPTQ, 262K): decode 66.6–70.0, prefill about 2,930–2,972, no measurable difference between versions (Strata #4996)実測2026-10-27リンク
Qwen3.8-Flash-NextStrataRTX 5060 Ti 16GB (PCIe 4 x8) + i7-14700 + 128GB RAM67Strata 0.2.7 (engine 0.1.52), standard-v1, own tuning: short 67.0 / 4K 63.8 / 16K 62.7; prefill 16K 1,834実測2026-10-28リンク
Bonsai2 27B未指定RTX 5060 Ti 16GB6767 tok/s with MTP head; 262K context + MTP + vision loaded together use 15.2/15.9 GB実測2026-09-27リンク
Qwen3.8-27BExLlamaRTX 5060 Ti 16GB66MiaAI_lab one-click package default config: code 66 tok/s, up to 176K context (temp 0, fixed seed, 3 runs consistent) (code speed; post gives no prose speed)実測2026-10-13リンク
Clover-35B-A3Bllama.cpp1× RTX 5060 Ti 16GB61.9Diffusion LLM, plain autoregressive (author self-spec crashes with expert offload on 16GB). Code 61.7 / prose 61.9; 8K prefill 992; 32K first token 39.9s; 128K 179.9s; peak VRAM 14.8 GiB. llama.cpp PR build cb09d4355; 512 tokens, temp 0, seed 42, thinking off, median of 2. Reference: same author Spark BF16/SGLang prose 26.0 (x-260). tok_s uses prose 61.9実測2026-11-06リンク
Qwen3.8-Flash-NextStrataRTX 5070 12GB + RTX 5060 Ti 16GB + 64GB RAM (WSL)54Strata v0.2.6 / v0.2.8, the same long text read 3 times: generation 54–62 (first, cold-cache run 41–55); prefill at about 4K on v0.2.8 221.7 / 289.6 / 333.5. The tester warns that the first run measures the file cache, not the version. On 2026-11-08 the same author's two-card setup, Strata 0.2.11, IQ3_S and 40960 limit uses STRATA_PREFILL_LAYER_MAJOR=1: 14K prefill 466.6 to 1059.8, 33K 453.7 to 717.5, medians of 3 (x.com/murasametech/status/2119247324656623651). Prefill only; no decode-speedup claim. Keep 54. On 2026-11-10 the same author, same two-card IQ3_S setup, compared v0.2.11 and v0.2.12 on the same day: decode 53.9–59.9 with no gap. Prefill only moved on the 128K-limit ~33K-token case, 411.7 to 496.2 (about +20%); other prefill ranges overlap. Keep headline 54.実測2026-11-02リンク
Qwen3.8-27Bllama.cppRTX 5060 Ti 16GB50.2MTP-2: 30-token prompt 50.2 / 6.6K prompt 48.1 tok/s; MTP-1 45.8; no MTP 27.6; peak VRAM 14.8 GiB; acceptance rate 77.2%実測2026-10-11リンク
Qwen3.8-Flash-NextStrataRTX 3060 12GB + RTX 5060 Ti 16GB + 125GB RAM39Strata 0.2.10 source build. First Ampere+Blackwell NVFP4 layer-split report: decode 39–45 (UD-Q4_K_XL on the same pair was 29.5–37.1); 31K prompt 24.8–25.2s (was 29.0s). Best split K=18. tok_s uses the low end 39実測2026-11-05リンク
Qwen3.8-27B-SLQ-EAR98vLLM2× RTX 5060 Ti 16GB33code 33.0, prose 33.0; aggregate (separate) 8 requests 143.1; context 94,208実測2026-10-26リンク
Qwen3.8-27B (Saluki by Underdog, 7.9 GB)llama.cppRTX 5060 Ti 16GB32code 32 tok/s, can fill 262K context (no vision file) (code speed; post gives no prose speed)実測2026-10-13リンク
Qwen3.8-27B未指定RTX 5060 Ti 16GB (Thunderbolt enclosure) + MacBook Pro M1 Max (Linux)24Decode 24, 2.6× the laptop GPU実測2026-11-03リンク
Qwen3.8-27Bllama.cpp2 cards: RTX 5060 Ti 16GB + GTX 1080 8GB (tensor split 4,1)21.9llama.cpp log 21.39–21.93 tok/s (generating 1.6–1.9K tokens); tg_3s ≈26.4実測2026-10-14リンク
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.実測2026-11-10リンク
Qwen3.8-Flash-NextBasaltRTX 5090 32GB + RTX 5060 Ti 16GB + 128GB RAM300Basalt 0.2.0: prose about 300, structured about 600, prefill about 10K. NVFP4-DQ8 KLD against BF16 is 0.0561主張2026-11-03リンク
Qwen3.8-27BtqLLMRTX 5060 Ti 16GB175.9Short prompt 175.9 tok/s (median of 5 runs, run locally, not submitted to LocalMaxxing); 200K prompt 113 tok/s公式2026-10-15リンク
Qwen3.6-35B-A3B-Escha-W2SGLangRTX 5060 Ti 16GB128128 tok/s decode主張2026-07-22リンク
Qwen3.8-Flash-NextStrata2× RTX 5060 Ti 16GB (layer split, main + helper) + i5-12600KF + 62.6GiB (Ubuntu 26.04)79.5Strata commit 628f7495, engine 0.1.59, release build; --draft dflash2 --spec 6, KV int8, layer split auto. Configured 204,800, 1 slot. Complete strata bench run standard (thinking on, temperature 1.0): headline is short decode median 79.5 (76.8-79.9, 3 runs). 4K 66.4, 16K 72.8. Prompt read 4K/16K: 838/1,600. Sanity 12/12. The body is the runner report, but results.json and outputs.jsonl are not attached to the issue, so this is a claim. The 2× 5060 Ti IQ2_XS row is comm-194.主張2026-11-13リンク
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 branch主張2026-10-21リンク
Qwen3.8-Flash-NextStrata1× RTX 5060 Ti 16GB36.5Strata v0.2.11: Dynamic v3.1 vs prior UD-IQ4_XS, chat 30.9→36.5 (+18%), agent 40.8→43.7; prefill almost unchanged. Note: coexists with Strata #5296 (gfx1151 prefill regression on the new files) — backends are not comparable. tok_s uses the new-file chat 36.5. Another tester jkenner111 uses 5060 Ti / 7900X / 64GB, source 0.1.58, UD-IQ4_XS V3.1 rev d9620119 and 262144 limit: standard short/4K/16K 56.9/52.0/48.3, sanity 12/12, about 9.4GiB experts still read from NVMe (github.com/Niko1221/Strata/pull/5514). Not a matched repeat of the original author/default setup; keep 36.5.主張2026-11-06リンク
Qwen3.8-27B-NVFP4-MTPllama.cppRTX 5070 Ti 16GB + RTX 5060 Ti 16GB (dual card tensor split)18.4418.44 tok/s decode; 645.6 tok/s prefill; MID-HIGH (Q8_0) + MTP at 32K 25.98 tok/s decode, 1451.6 tok/s prefill主張2026-08-14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。