X local-model bench digest — Nov 3, 2026
As of Nov 3, 2026 — 35 posts collected from X. 122 measured runs catalogued, 35 kept to the digest only. Latest source post Nov 3, 2026 09:39 UTC.
Every number below is what its author published, transcribed with the engine, quantization and context length they reported. Sources are linked on each entry. Only posts that state the full methodology (engine build, quantization, context length, hardware state, measurement method) reach the benchmark board; the rest stay here as leads.
RTX 5090 × Qwen3.8-27B
Nov 3, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 1220.0 tok/s | — |
💥 一张 RTX 5090，Qwen3.8-27B 单路解码冲到 1220 tok/s。 日本开发者 YASAKANI 自测：定格无超频，同机比 MegaCapybara 大约快 1.9 倍。 温馨提示：作者自测，第三方复现和开源都还没来。 你更信视频，还是等复现？ 来源：@yasakani_dev👇 https://t.co/e3WIxdAoyg
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Nov 3, 2026 · by @bebuu_ai · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 202.0 tok/s | — |
DeepSeek Harness has a "Creator Mode" where the harness can modify its own source code while you are using it. A developer asked it to add Claude Code account support, and DeepSeek V4.1 Flash (running at 202 tok/s with a 99% cache hit) wrote the OAuth plugin, patched its own config, ran smoke tests, and hot-loaded the feature into itself in 17 minutes. Now he can use his claude account into DSH.
Digest-only: confidence 0.55 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Nov 3, 2026 · by @rlaope
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 1000.0 tok/s | — |
I got DeepSeek V4.1 Flash running at ~1000 tok/s. Same open weights, served on our own GPUs. That's ~3x faster than DeepSeek's own API. Why it matters: - coding agents stop making you wait between steps - a 3k-token answer lands in a few seconds, not ten - long agent loops and batch jobs finish in a fraction of the time Same model, same quality. Just faster. Then I opened it up so you can use it. OpenGateway ›…
Digest-only: confidence 0.40 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 5090 × Qwen3.8-27B
Nov 3, 2026 · by @yasakani_dev
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | yasakani | 1220.0 tok/s | — |
| — | — | yasakani | 937.0 tok/s | — |
| — | — | yasakani | 424.0 tok/s | — |
| — | — | yasakani | 532.0 tok/s | — |
| — | — | yasakani | 1220.0 tok/s | — |
| — | — | yasakani | 424.0 tok/s | — |
| — | — | yasakani | 613.0 tok/s | — |
| — | — | yasakani | 745.0 tok/s | — |
| — | — | yasakani | 1241.0 tok/s | — |
| 8192 | — | yasakani | 1260.0 tok/s | — |
| — | — | yasakani | 381.0 tok/s | — |
| 250K | — | yasakani | 152.0 tok/s | — |
YASAKANIエンジンで、RTX 5090 の単一デコード 1,220 tok/s を達成!! 今回は世界最速のエンジンである証明をしました。 RTX 5090 1枚・Qwen3.8-27B・1本の生成で、公開されている最速記録を5本すべて上回った（定格・同じ土俵の条件で、最大2.5倍）。 「5090で最速」を名乗る MegaCapybara 1.39 とは、同じPC・定格・同じ問題で直接対決。動画は実際の生成を、記録した時刻どおりに再生しています。 ■ 公開記録 → YASAKANI（tok/s） - mkyt（JSON）500 → 1,220（2.44倍） - NInfer（構造化3問）377 → 937（2.49倍） - TensorFold（英語）298 → 424（1.42倍） - ignis（コード）491 → 532（1.08倍） - MegaCapybara（コード）540※ →…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization).
Arc A750 × Qwen3.8
Nov 3, 2026 · by @QwenAIFans · third-party report
💥 8GB 显卡，终于能把 Qwen3.8 开到 128K 了。 Strata 0.2.10：模拟 8GB 跑 DFlash2，解码约 63 到 73 tok/s；真机 Arc A750 大约 12.7 到 14.8。默认改走更小的 DFlash2，不再硬塞 MTP。 温馨提示：系统内存还是要够，官方测的是大内存机器。 来源：GitHub Strata👇 https://github.com/Niko1221/Strata/releases/tag/v0.2.10 https://t.co/7pdXxeqi8a
Digest-only: confidence 0.55 below 0.60; incomplete methodology (quantization, hardware state, measurement method).
Intel Arc A750 × Qwen3.8-Flash-Next
Nov 3, 2026 · by @NFT_Chen · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 128K | — | strata v0.2.10 | 20.0 tok/s | — |
| — | — | strata v0.2.10 | 399.0 tok/s | — |
🤯炸了！仅8GB显卡也能畅玩125B级Qwen3.8 Flash Next，速度稳定20+ tok/s！支持128K上下文！多卡并行暴涨近100%，DFlash2轻量drafter让小显卡也能飞！ Strata v0.2.10正式发布，把本该需要几十上百GB显存的Qwen3.8-Flash-Next（125B MoE，每token仅激活6B）硬生生压进8GB显卡，还能稳住20+ tok/s！ 核心亮点 🔹8GB显卡就能跑：动态内存管理 + 专家卸载，128K上下文不再是梦，实测Intel Arc A750等卡片稳定出字。 🔹速度起飞：DFlash2替代MTP（仅约0.7GB），小显卡解码更快；双卡最高+98%（2×RTX A6000从201→399 tok/s），4×R9700 +78%。 🔹并行爆炸：All-Cards Split Engine默认开启，多请求吞吐大幅提升，长文档等待时间砍近一半。…
Digest-only: confidence 0.55 below 0.60; incomplete methodology (quantization, hardware state, measurement method).
DGX Spark × Qwen3.8 Flash Next
Nov 3, 2026 · by @plotarmordev
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | tensorfold | 75.0 tok/s | — |
| — | — | vllm | 95.0 tok/s | — |
| — | — | tensorfold | 114.0 tok/s | — |
| — | — | vllm | 145.0 tok/s | — |
| 8K | — | tensorfold | — | 2200.0 tok/s |
| 8K | — | vllm | — | 2956.0 tok/s |
The 128GB DGX Spark went from $3999 to $6950 this year, but the software got faster too! Qwen3.8 Flash Next on one Spark, TensorFold vs vLLM: 2 users 75 > 95 tok/s 4 users 114 > 145 tok/s 8K prefill 2200 > 2956 tok/s So 27% more decode speed and 34% faster prefill, is that https://t.co/OrR4ev6yQ3
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state, measurement method).
RTX 4090 × Apex-35B-v1.0
Nov 3, 2026 · by @servasyy_ai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata | 115.0 tok/s | — |
| — | — | strata | 205.0 tok/s | — |
重磅！兄弟们又有新模型发布了！ 我刚拿手头上的4090 硬肝了三十多个小时（满载跑了 17 小时，烧了400 万 token），把 @apexin_ai 刚发的 Apex-35B-v1.0 完整测试了一遍。 老规矩，先给结论：如果你想在单卡上跑一个真能自己干活的研究助手，超衍的这个模型可以说独一档，我最欣赏的是它的递归自我改进能力，写出来的论文和人类几乎相当！而且在很多方面还超越世界先进模型。 我在 48G 的 4090 上连测了九项，跟大家分享三个最重要的点： 1、用 Strata 跑，显存只给 24G 就能干到 115 tok/s，比 llama.cpp 快 72%；48G 开满直接飙到 205 tok/s。 2、丢给它一道 64 根血管的流体优化题，它在 30 轮里，从数据摸出流量和半径的 3.5 次方规律，把误差缩小了 580 倍，并最终完成了目标！ 3、用 Claude Code 接上它修一个挂掉的 Python…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
RTX 3060 × Qwen3.8-35B-A3B
Nov 3, 2026 · by @Oluwaphilemon1 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 262K | Q4_K_M | llama.cpp | 50.0 tok/s | — |
| 262K | Q4_K_M | llama.cpp | — | 550.0 tok/s |
A 12GB RTX 3060 is reportedly generating around 50 tokens/s with Qwen3.8-35B-A3B, a model with 35 billion parameters. The same setup reaches approximately 550 tokens/s during prompt processing. No 24GB GPU. No expensive multi-GPU configuration. Just a consumer RTX 3060, 16GB of system RAM and some careful inference optimisation. That’s an interesting result for anyone trying to run larger models locally. The…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
8GB device × Qwen3.8 Flash Next
Nov 3, 2026 · by @servasyy_ai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata 0.2.10 | 20.0 tok/s | — |
你敢想象吗？ 8G都能玩Qwen3.8 Flash Next了！而且速度到20 tok/s+ Strata 0.2.10大升级，没有玩过的朋友赶紧去试试吧！ https://t.co/Umt3fyDyjW
Digest-only: confidence 0.55 below 0.60; incomplete methodology (quantization, context length, hardware state, measurement method).
RTX A6000 x2 × Qwen3.8 Flash Next
Nov 3, 2026 · by @murasametech
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata v0.2.10 | 399.0 tok/s | — |
| — | — | strata v0.2.10 | 201.0 tok/s | — |
| — | — | strata v0.2.10 | 20.0 tok/s | — |
Strata v0.2.10 リリース！🚀 Qwen3.8 Flash Nextが8GBグラフィックカードで20+ tokens/sを実現。並列処理が大幅に進化し、全体的に大きな性能向上が入りました。 主なハイライト： ・2×NVIDIAカード：最大+98%高速化（2×RTX A6000で201→399 tok/s） ・4×Radeon AI PRO R9700：+78%高速化 ・長文ドキュメント：最初の応答までの待ち時間が約21秒→12秒（-44%） ・Strix Halo：+12% ・M5 Max：エージェント向け+10.3% ・DGX SparkやMacのprompt処理も改善 新機能・改善点： ・DFlash2がQwen3.8のデフォルトdrafterに（約0.7GBと軽量、MTP層を回避） ・Linuxで「All-Cards Split Engine」がデフォルトに（マルチGPUで大幅スループット向上）…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (quantization, context length, hardware state, measurement method).
RTX A6000 x2
Nov 3, 2026 · by @jun1228909
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata v0.2.10 | 399.0 tok/s | — |
| — | — | strata v0.2.10 | 350.0 tok/s | — |
？！ 予告なしで大型アプデが！🤯 Strata v0.2.10 リリース 🚀 今回はmulti-GPU split engine + DFlash2実用化 + DeepSeek強化、といったところか🧐 ・複数GPUをより並列に使う新しいsplit engineを追加 ・2×RTX A6000で編集系workload 201→399 tok/sの例 ・4×R9700でもsplit engineを自動選択、8並列で約350 tok/s ・DFlash2のAuto選択を強化、MTP / DFlash2 / split engineを環境に応じて切替 ・RTX PRO 6000 Blackwell 2枚で32K promptを約3.3〜3.5秒 ・Intel Arc A750 8GBでもAutoでDFlash2動作 ・DeepSeek Coder対応を追加 ・MacのDeepSeek Vision対応を強化 ・Strix Halo +…
Digest-only: model.name missing; incomplete methodology (quantization, hardware state, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 23 posts from the same window are listed here in short form, each with its source post.
- DGX Spark x4 × DeepSeek-V4.1-Flash-1017 — 230.7 tok/s decode (5 measurements) · @Tech2Wild
- M3 Ultra x2 × Mistral Large 4 — 20.0 tok/s decode (1 measurement) · @Prince_Canuma
- RTX 3080 × Groot 19B A3B — 52.0 tok/s decode (4 measurements) · @LiLensss
- RTX 5090 × Qwen3.8-Flash-Next — 127.1 tok/s decode (5 measurements) · @LiLensss
- DGX Spark — 146.8 tok/s decode (2 measurements) · @TandemLLM
- Strix Halo (Ryzen AI Max+ 395) + RX 7900 XTX x2 × DeepSeek sparse-attention indexer (Project Maya) — 82.5 tok/s decode (4 measurements) · @jgattermayer
- DGX Spark x8 × DeepSeek-V4.1-Flash-1017 — 287.0 tok/s decode (5 measurements) · @wei_wang
- RTX 3090 — 80.0 tok/s decode (2 measurements) · @VRAMCalculator
- Unspecified platform × Qwen — 1500.0 tok/s decode (2 measurements) · @net_termina
- RTX PRO 6000 x2 × GLM-5.3-Flash — 436.0 tok/s decode (7 measurements) · @wei_wang
- M5 Mac Studio 256GB × DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-3.0bpw — 100.0 tok/s decode (1 measurement) · @rez0__
- Mac Studio M3 Ultra 512GB × DeepSeek V4.1 Flash — 85.0 tok/s decode (3 measurements) · @Jhaddix
- H200 × deepseek — no throughput stated · @0xSero
- DGX Spark × Qwen 3.8 Flash Next — 70.0 tok/s decode (1 measurement) · @twid
- Unspecified platform × deepseek 1017 abliterated — 285.0 tok/s decode (5 measurements) · @aaronjmars
- RTX PRO 6000 × Qwen3.8-Flash-Next — 503.0 tok/s decode (4 measurements) · @0x0SojalSec
- CMP 170HX x4 × DeepSeek V4.1 Flash 1017 — 765.0 tok/s decode (7 measurements) · @OwnYourCompute
- Unspecified platform × DeepSeek V4.1-Flash GSQ uncensored — 50.0 tok/s decode (1 measurement) · @rS_alonewolf
- DGX Spark x2 × GLM-5.3-Flash — 83.0 tok/s decode (12 measurements) · @atasrb
- DGX Spark x4 × DeepSeek-V4.1-Flash-1017 — 144.6 tok/s decode (10 measurements) · @niklaslenz_ai
- Mac × Qwen3.8-Flash-Next — 103.0 tok/s decode (1 measurement) · @QwenAIFans
- RTX 5070 12GB + RTX 5060 Ti 16GB × Qwen3.8-Flash-Next — 353.0 tok/s prefill (6 measurements) · @murasametech
- NVIDIA GB300 x8 × DeepSeek V4.1 Flash — 1567.0 tok/s decode (1 measurement) · @Schnitzel
How to read this page
This digest is an automated transcription of public posts. Runs with a stated engine build, quantization, context length, hardware state and measurement method are published straight to the benchmark board with a link to their source; the rest stay here. These are author-reported numbers, not an independent DiggTop re-measurement: every row carries its provenance so you can check the original, and a number that turns out to be misreported is corrected or removed. Our own testing follows the methodology. Spotted an error? Reply on the source post or submit a corrected run at /submit.
Sources
- @QwenAIFans — Nov 3, 2026
- @bebuu_ai — Nov 3, 2026
- @rlaope — Nov 3, 2026
- @yasakani_dev — Nov 3, 2026
- @QwenAIFans — Nov 3, 2026
- @NFT_Chen — Nov 3, 2026
- @plotarmordev — Nov 3, 2026
- @servasyy_ai — Nov 3, 2026
- @Oluwaphilemon1 — Nov 3, 2026
- @servasyy_ai — Nov 3, 2026
- @murasametech — Nov 3, 2026
- @jun1228909 — Nov 3, 2026
- @Tech2Wild — Nov 2, 2026
- @Prince_Canuma — Nov 2, 2026
- @LiLensss — Nov 2, 2026
- @LiLensss — Nov 2, 2026
- @TandemLLM — Nov 2, 2026
- @jgattermayer — Nov 2, 2026
- @wei_wang — Nov 2, 2026
- @VRAMCalculator — Nov 2, 2026
- @net_termina — Nov 2, 2026
- @wei_wang — Nov 2, 2026
- @rez0__ — Nov 2, 2026
- @Jhaddix — Nov 2, 2026
- @0xSero — Nov 2, 2026
- @twid — Nov 2, 2026
- @aaronjmars — Nov 2, 2026
- @0x0SojalSec — Nov 2, 2026
- @OwnYourCompute — Nov 2, 2026
- @rS_alonewolf — Nov 2, 2026
- @atasrb — Nov 2, 2026
- @niklaslenz_ai — Nov 2, 2026
- @QwenAIFans — Nov 2, 2026
- @murasametech — Nov 2, 2026
- @Schnitzel — Nov 2, 2026
FAQ
Are the numbers in this digest sourced?
They are transcribed from public X posts and the repositories those posts link to, with attribution and a source link on every entry. Rows that state the full methodology are published to the benchmark board automatically, so treat them as author-reported until a dispute is raised: the source link on every row is the check.
What does it take for a row to reach the sourced tier?
Engine and exact version, quantization, context length, hardware state and a described measurement method. Posts that state all five are published to the benchmark board as soon as the collector reads them, each row carrying its source link — nothing waits in a human review queue. Anything missing a condition stays in this digest.
Where does the data come from?
Public X posts about local-model inference measured on owned hardware, discovered by keyword search every two hours and linked back to the original thread. The digest is regenerated daily.