X local-model bench digest — Oct 18, 2026
As of Oct 18, 2026 — 10 posts collected from X. 34 measured runs catalogued, 2 published to the board, 8 kept to the digest only. Latest source post Oct 17, 2026 19:30 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.
DGX Spark × Qwen3.8 Flash Next
Oct 17, 2026 · by @MiaAI_lab
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 256K | — | tensorfold | 46.0 tok/s | — |
| 256K | — | tensorfold | 132.0 tok/s | — |
| 256K | — | tensorfold | — | 2000.0 tok/s |
Rejoice single DGX Spark owners! You can now run Qwen3.8 Flash Next on TensorFold with the new Nvidia FP4 QAD quant 🔥 - 1.1M total FP8 KV cache - 256k context, 5 concurrent users - 46-55 tok/s on prose, single stream - 132 tok/s on prose, 5 concurrent steams - Prefill 2000-2300 tok/s Qwen3.8 Flash Next is still the BEST model to run on one DGX Spark, bare none! You can choose what to run: FP4 QAD for quality, or…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
DGX Spark × Qwen3.8-Flash-Next
Oct 17, 2026 · by @ViC305
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | FP4-Local | — | 68.0 tok/s | — |
🚨 @nvidiaai FP4-LOCAL: THE GOOD, THE BAD & THE UGLY! 🔥 I tested Qwen3.8-Flash-Next 4 ways. FP4-Local TP1 & 2, NVFP4 and EXL3 FP4: 1 Spark: 90.8% SixCat, 68 tok/s. 2 Sparks: 32% slower decode. DIY quantization? That's the ugly part. My full article 👇 https://x.com/i/article/2111519852049625088
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
PS5 × Qwen 27B
Oct 17, 2026 · by @hkepcmedia
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | llama.cpp | 20.0 tok/s | — |
| — | — | llama.cpp | 61.0 tok/s | — |
| — | — | llama.cpp | 124.0 tok/s | — |
【用來玩 AI 😱】Sony PS5 遊戲越獄（Jailbreak）生態迎來突破性進展。開發者社群近期成功將知名推論框架 llama.cpp 改編為原生應用程式並移植至 PS5，直接調用 PS5 內置的 AMD GPU 硬體加速與高達 11.4 GiB 的系統記憶體。經實際測試，越獄 PS5 在本機離線運行大型語言模型（LLM）的生成速度最高可突破每秒 124 個 Token（Tokens/s），甚至能順暢運算 270 億參數（27B）的中大型 AI 模型。 https://t.co/Or8rErK4wf 回顧此前進展，GitHub 開發者 cobanov 於兩週前發佈了初步移植至 PS5 的「PS5LM」，但該版本僅以 CPU Payload 方式執行，實質上是在主畫面與系統背景服務運作的同時插入指令。受限於系統防護機制，一旦佔用過多記憶體便會導致主畫面凍結崩潰，可用記憶體容量因而被鎖死在約 6.25 GiB…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
DGX Spark x4 × DeepSeek-V4.1-Flash-1017
Oct 17, 2026 · by @Tech2Wild
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 300K | MXFP4+MXFP8 | vllm dsv41-feat (commit e47aa780b) | 165.7 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat (commit e47aa780b) | 76.3 tok/s | — |
| 93335 | MXFP4+MXFP8 | vllm dsv41-feat (commit e47aa780b) | — | 1903.0 tok/s |
I ran my own benchmark on DeepSeek V4.1 Flash 1017 and here's what it scored on 4x DGX Spark 🔥 ⚡ 165.7 tok/s throughput 🎯 98.6 on my 69-scenario tool eval Full recipe 👇 https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark https://t.co/k0v54w6kJj
Full methodology stated (engine, quantization, context, hardware state, method) — 1 added to the benchmark board (1 of 3 measured rows queued; the rest are published here only).
Unspecified platform × deepseek v4.1 flash
Oct 17, 2026 · by @MingTian685474 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 300.0 tok/s | — |
看到有不少人模型指纹测出来 deepseek v4.1 flash 1017 为 gpt-6-Astra 然后就有了两种观点： 1. deepseek 把 v4.1 flash 路由到 Astra 了。但是你有没有想过，你见过 300 tok/s 的 Astra？？？那 OpenAI 都得过来学学怎么做的 2. deepseek 把 Astra 给蒸馏了。有没有蒸馏我不知道，但是通过这个模型指纹测出来的结果去判断蒸馏就是不合理的 像这类模型指纹工具，原理主要是让模型生成特定区间的随机数，再去和库里已知模型的分布做特征匹配 也就说这里如果压根没有收录 v4.1 flash 的特征（就算收录过旧版 v4.1 flash 的特征，后训练一改权重，随机数分布也会改变），那算法就只会去现有库里找一个统计距离最为相近的模型强行匹配上去。匹配出 Astra 只能说明在现有样本里它俩距离最近，根本不代表分布真的一样…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark x4 × DeepSeek-V4.1-Flash-1017
Oct 17, 2026 · by @Tech2Wild
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 76.3 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 75.8 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 63.2 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 56.2 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 55.6 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 36.1 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 35.8 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 31.7 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 165.7 tok/s | — |
DeepSeek V4.1 Flash 1017 updated numbers 1 stream tok/s, old → new: 💻 Code 73.8 → 76.3 📊 Format 55.4 → 75.8 🧾 JSON 52.1 → 63.2 ➗ Math 50.9 → 56.2 🧠 Reasoning 37.8 → 55.6 ✍️ Prose 24.4 → 36.1 📝 Summary 25.7 → 35.8 📖 Narrative 24.9 → 31.7 https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark
Full methodology stated.
DGX Spark x4 × DeepSeek-V4.1-Flash-1017
Oct 17, 2026 · by @Tech2Wild
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 165.7 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 76.3 tok/s | — |
| 93335 | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | — | 2201.0 tok/s |
| 300K | EXL3 3.5bpw | vllm bot-lab-21's EXL3 build + drowzeys' abliteration… | 190.0 tok/s | — |
| 300K | EXL3 3.5bpw | vllm bot-lab-21's EXL3 build + drowzeys' abliteration… | 152.9 tok/s | — |
| 300K | MXFP4+MXFP8 | vllm dsv41-feat overlay, commit e47aa780b | 131.9 tok/s | — |
DeepSeek V4.1 Flash 1017 on 4x DGX Spark (TP4) 🚀 Same 4 Sparks vs the old V4.1 Flash: ⚡ 6 streams: 131.9 → 165.7 tok/s throughput (+26%) 📥 93K token prompt: 78s → 42s 🧠 300K context, KV pool 1.07M → 1.52M tokens Per prompt numbers 👇 https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/tree/main/runs/2026-10-17-1017
Full methodology stated (engine, quantization, context, hardware state, method) — 2 added to the benchmark board (2 of 6 measured rows queued; the rest are published here only).
RTX 5090 × Qwen
Oct 17, 2026 · by @MrHydeDev · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 262K | — | llama.cpp | 90.0 tok/s | — |
| 262K | — | llama.cpp | 55.0 tok/s | — |
Sigo dándole herramientas de hacking a mi Qwen abliterated para que intente romper nuestras propias aplicaciones y no va mal, la verdad Le paso una URL y unas instrucciones con lo que quiero revisar, y a partir de ahí le dejo investigar. Va probando cosas, leyendo los resultados y decidiendo por dónde seguir. Si encuentra una pista, tira de ella (y creeme que tira hasta agotar todas las posibilidades, a veces se…
Digest-only: confidence 0.35 below 0.60; incomplete methodology (engine+version, quantization, hardware state, measurement method).
RTX 5090 × Llama 3.1 8B
Oct 17, 2026 · by @vibelancer · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | FP16 | — | 69.0 tok/s | — |
| — | FP16 | — | — | 7029.0 tok/s |
| — | FP16 | — | 38.0 tok/s | — |
| — | FP16 | — | 383.0 tok/s | — |
| — | FP16 | — | 42.0 tok/s | — |
same 8B model. 69 tok/s on an RTX 5090, 63 on a Mac Studio M3 Ultra. then you paste 12,500 tokens and one starts answering in 2s, the other in 12s llmspeed is a static site that streams an answer at the tok/s your rig would really hit, after the real wait for the first token. 30 GPUs, Macs and CPUs × 15 models · 1,446 combos. no backend. no signup. Vite + Preact, MIT, one json of predictions most people still buy…
Digest-only: confidence 0.35 below 0.60; incomplete methodology (engine+version, context length, hardware state, measurement method).
Unspecified platform × Qwen3.8-27B
Oct 17, 2026 · by @0xBakeer
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 150K | EXL3 6.0 | tandemllm | 35.0 tok/s | — |
I gave Qwen3.8-27B EXL3 6.0 one task in TandemLLM: build a city of pagoda gardens and keep working until I say stop. 2 hours later it's a walled voxel city with tall pagodas, cherry trees, glowing lanterns, little boats on the river, and its own city controls panel. I started it from my phone and walked away. No need to watch. It implements, tests and takes screenshots, and a decision model picks what to do next…
Digest-only: hardware.name missing; incomplete methodology (engine+version, hardware state, measurement method).
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
- @MiaAI_lab — Oct 17, 2026
- @ViC305 — Oct 17, 2026
- @hkepcmedia — Oct 17, 2026
- @Tech2Wild — Oct 17, 2026
- @MingTian685474 — Oct 17, 2026
- @Tech2Wild — Oct 17, 2026
- @Tech2Wild — Oct 17, 2026
- @MrHydeDev — Oct 17, 2026
- @vibelancer — Oct 17, 2026
- @0xBakeer — Oct 17, 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.