X local-model bench digest — Oct 17, 2026
As of Oct 17, 2026 — 29 posts collected from X. 67 measured runs catalogued, 29 kept to the digest only. Latest source post Oct 17, 2026 11:18 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.
Unspecified platform × DeepSeek-V4.1-Flash
Oct 17, 2026 · by @Zekunnnnn · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 230.0 tok/s | — |
DeepSeek-V4.1-Flash got faster with the 1017 update. It's at ~230 tok/s on DeepSeek's own API and even faster than Haiku 5.5 (~210). My own tracking. Hurry up and scale the model. Huawei, ramp up Ascend 950DT supply ASAP https://t.co/ArFpyg1C7D
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash 1017
Oct 17, 2026 · by @runinfrai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 1M | — | — | 327.0 tok/s | — |
DeepSeek V4.1 Flash 1017 is live on RunInfra hour 0 support. day 0 is for slow teams 327 tok/s output. 30ms to first reasoning token $0.15 per 1M input, $0.60 per 1M output, $0.01 per 1M cached. 1M token context cache hit rate stays above 98% on agentic workloads. every hit and miss shows up in your dashboard, so you can check it yourself OpenAI-compatible chat completions and Anthropic-compatible /v1/messages. text…
Digest-only: hardware.name missing; incomplete methodology (engine+version, quantization, hardware state, measurement method).
Unspecified platform × DeepSeek-V4.1-Flash
Oct 17, 2026 · by @vllm_project · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 830K | — | vllm | 100.0 tok/s | — |
🎉 Congrats @deepseek_ai on DeepSeek-V4.1-Flash-1017! Better scores and shorter outputs make it a great fit for long agent loops. Making each step fast is the serving side's job. On DeepSeek-V4.1-Flash, vLLM delivers 1.3B tokens per dollar TCO on SemiAnalysis AgentX: 830K tok/s per GPU (cached + input + output) at 100 tok/s per user. 🚀 It's been a fun few weeks building with the community. Chapter one, 5x agentic…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Oct 17, 2026 · by @SPAC89 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 244.0 tok/s | — |
| — | — | — | 300.0 tok/s | — |
DeepSeek V4.1 Flash 1017 is hitting 244 tokens per second, that's incredible! It's even faster than GPT 6.1 Sol in Ultrafast mode while offering intelligence comparable to GPT 5.6 Sol, what a time to be following AI! I still can't understand how they can offer this level of quality and speed at such a low price, and from my first tests I've seen speeds reaching almost 300 tokens per second. It now has its own built…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × deepseek v4.1 flash
Oct 17, 2026 · by @vorpal_onchain · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 60.0 tok/s | — |
you don't need a jailbreak for deepseek v4.1 flash anymore xplabs has an uncensored version, and it's free 😳 - 1.05M context - 60 tok/s, tested try it here : https://platform.xplabs.ai https://t.co/VApd1c2S4P
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Oct 17, 2026 · by @Fei2411 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 250.0 tok/s | — |
这次真无敌了😭DeepSeek V4.1 Flash-1017 正式上线！官宣：“许愿式编程一键成真”，基准数据甚至大幅度优于 GPT 5.6 Sol ！ 在我的测试中，DeepSeek V4.1 Flash-1017 有非常棒的设计品味，我也看到了非常多优秀的 3D 作品展示，并且社区也有反馈说这个模型语言表现好了很多。总之这是个非常优秀的模型！ 甚至，它依然是个 Flash 模型，在我的使用过程中，几乎全程保持 250 tok/s 的高速！！！
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Oct 17, 2026 · by @painn_x · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 60.0 tok/s | — |
DeepSeek V4.1 Flash UNCENSORED is FREEE 😳 no jailbreak needed, straight up uncensored what you get: - DeepSeek V4.1 Flash uncensored - 1.05M context - 60 tok/s, tested try it here: https://platform.xplabs.ai https://t.co/6cWgMngMBn
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
CMP 170HX x4 × DeepSeek-V4.1-Flash-EXL3-3.0bpw
Oct 17, 2026 · by @BBacktesting
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | EXL3 3.0bpw | vllm | 178.3 tok/s | — |
| — | EXL3 3.0bpw | vllm | 587.0 tok/s | — |
| 2048 | EXL3 3.0bpw | vllm | — | 1570.0 tok/s |
DeepSeek V4.1 Flash 4× CMP 170HX optimized drop 🎙️⚡️ 178 tok/s, one user. 587 tok/s across 8. 🔥 Prefill 1,570 tok/s at 2K. KV cache 617K tokens. EXL3. Full graphs and coop underneath, tensor-core MoE from row 43 up. @Morrowmake https://github.com/CurtisAccelerate/DeepSeekv41-Flash-EXL3-TC-MoE
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version).
RTX PRO 6000 × Qwen3.8-Flash-Next
Oct 17, 2026 · by @wei_wang
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 32768 | FP4 | — | 33.0 tok/s | — |
| — | FP4 | — | 49.0 tok/s | — |
| — | FP4 | — | — | 2100.0 tok/s |
NVIDIA 给 Qwen3.8-Flash-Next 出了一个官方本地版，FP4 权重 96 GiB，专门在 RTX PRO 6000 和 DGX Spark 上验过。 精度只掉了一点。GPQA Diamond 91.5%，BF16 原版 92.2%。ToolEvalBench 88.7%，原版 89.4%。 单张 PRO 6000 要把 PLE 那张 n-gram 表放到内存里，FP8 KV 能放 85 万 token。 单台 Spark 的一组实测：decode 从 49 tok/s 起步，32K 上下文时 33，prefill 一直在 2.1K 左右。 我那张 PRO 6000 跑 Flash-Next，用的是自己写的 NVFP4 KV 补丁，KV 池 150 万 token。这版官方权重我准备换上去，和现在的配置对比一下。
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
RTX 3060 × qwen
Oct 17, 2026 · by @sudoingX · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 262K | — | — | 50.8 tok/s | — |
i keep watching people walk right past the door and the door to local ai is a 12, 16 or 24gb gaming gpu the rtx 3060 is still the #1 desktop gpu on steam 16gb is now the most common vram on steam, 26.9% of everyone 65.7% of steam sits between 8 and 16gb that's the hardware. here's what it runs today, measured: 8gb: 27b with a 48k ctx 12gb: that same 27b at 50.8 tok/s with 131k ctx 16gb: 262k ctx, and it sees 24gb…
Digest-only: confidence 0.35 below 0.60; incomplete methodology (engine+version, quantization, hardware state, measurement method).
RTX 5070 12GB + RTX 5060 Ti 16GB × Qwen3.8-Flash-Next
Oct 17, 2026 · by @murasametech
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 4K | IQ3_XXS | strata v0.1.43 | — | 728.0 tok/s |
| 14K | IQ3_XXS | strata v0.1.43 | — | 1315.0 tok/s |
| 32K | IQ3_XXS | strata v0.1.43 | — | 1404.0 tok/s |
2枚のGPU（RTX 5070 12GB＋RTX 5060 Ti 16GB）で、Strata v0.1.43 と Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS をresident方式で動かしました。 モデルは合計75.84GBで、前回のIQ2_XS（68.03GB）より一回り大きいです。 長文の読み込み速度（prefill）は、4K tokで約728、1.4万tokで約1,315、3.2万tokで約1,404 tok/sでした。 以前IQ3_Sで起きた「長文のたびにSSDから約88GBを読み直して約150 tok/sまで落ちる」現象は起きず、読み込みは1回あたり22〜121MBで済みました。 IQ2_XSと比べると、prefillは最大2割ほど、短文の生成は81〜88 tok/s（IQ2_XSは95〜98）と少し遅めです。…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (hardware state, measurement method).
RTX 3090 × Qwen3.8-Flash-Next
Oct 17, 2026 · by @Oluwaphilemon1
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 261K | IQ3_S | strata | 66.0 tok/s | — |
| — | IQ3_S | strata | 70.0 tok/s | — |
| — | IQ2_XS | strata | 87.0 tok/s | — |
| — | — | strata | — | 1900.0 tok/s |
If you own a single RTX 3090 with 24GB VRAM, Qwen3.8-Flash-Next is worth a look. This is a 177B-parameter model running through Strata by splitting its workload between GPU VRAM and system RAM. The GPU keeps the experts it uses most readily available, while the remaining model data stays in system memory. Here are the reported numbers from a used RTX 3090: IQ3_S Around 70 tokens/sec Still around 66 tokens/sec with…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 17 posts from the same window are listed here in short form, each with its source post.
- DGX Spark x2 × GLM-5.3-Flash — 62.2 tok/s decode (1 measurement) · @wei_wang
- MI355X x8 × Kimi_Moonshot K3 — 7567.0 tok/s decode (2 measurements) · @vllm_project
- A100 × llama.cpp — 93.0 tok/s decode (1 measurement) · @fahdmirza
- Unspecified platform × GLM-5.3 — 800.0 tok/s decode (4 measurements) · @ItsmeAjayKV
- DGX Spark × Multiverse Comet 60B — 56.2 tok/s decode (4 measurements) · @niklaslenz_ai
- DGX Spark × Qwen3.8-27B — 115.0 tok/s decode (6 measurements) · @FT_IOxCS
- DGX Spark × Nemotron 3.5 Bolt — 195.1 tok/s decode (2 measurements) · @ViC305
- DGX Spark × llama.cpp — 42.1 tok/s decode (1 measurement) · @sudoingX
- Tesla V100 x2 × GLM 5.3 Flash — 54.5 tok/s decode (1 measurement) · @PeasantSmith
- DGX Spark × Nemotron 3.5 Bolt 74B — 179.0 tok/s decode (5 measurements) · @niklaslenz_ai
- DGX Spark (GB10) x4 × GLM-5.3-Flash-NVFP4 — 199.0 tok/s decode (7 measurements) · @mmastrac
- GB300 × DeepSeek-V4.1-Flash — 30.0 tok/s decode (1 measurement) · @steeve
- RTX 5060 Ti × Nemotron 3.5 Bolt 74B — 309.8 tok/s decode (4 measurements) · @niklaslenz_ai
- RTX 3060 — 50.8 tok/s decode (2 measurements) · @sudoingX
- DGX Spark × Nemotron 3.5 Bolt — 166.0 tok/s decode (2 measurements) · @MiaAI_lab
- DGX Spark × Nemotron 3.5 Super VL — 38.0 tok/s decode (1 measurement) · @WescheNex1q
- Unspecified platform — 874.0 tok/s decode (1 measurement) · @sophiamyang
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
- @Zekunnnnn — Oct 17, 2026
- @runinfrai — Oct 17, 2026
- @vllm_project — Oct 17, 2026
- @SPAC89 — Oct 17, 2026
- @vorpal_onchain — Oct 17, 2026
- @Fei2411 — Oct 17, 2026
- @painn_x — Oct 17, 2026
- @BBacktesting — Oct 17, 2026
- @wei_wang — Oct 17, 2026
- @sudoingX — Oct 17, 2026
- @murasametech — Oct 17, 2026
- @Oluwaphilemon1 — Oct 17, 2026
- @wei_wang — Oct 16, 2026
- @vllm_project — Oct 16, 2026
- @fahdmirza — Oct 16, 2026
- @ItsmeAjayKV — Oct 16, 2026
- @niklaslenz_ai — Oct 16, 2026
- @FT_IOxCS — Oct 16, 2026
- @ViC305 — Oct 16, 2026
- @sudoingX — Oct 16, 2026
- @PeasantSmith — Oct 16, 2026
- @niklaslenz_ai — Oct 16, 2026
- @mmastrac — Oct 16, 2026
- @steeve — Oct 16, 2026
- @niklaslenz_ai — Oct 16, 2026
- @sudoingX — Oct 16, 2026
- @MiaAI_lab — Oct 16, 2026
- @WescheNex1q — Oct 16, 2026
- @sophiamyang — Oct 16, 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.