DiggTop
Menu
⌕

X local-model bench digest — Oct 22, 2026

As of Oct 22, 2026 — 37 posts collected from X. 78 measured runs catalogued, 37 kept to the digest only. Latest source post Oct 22, 2026 11:04 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.

Strix Halo 395+ 128GB × GLM 5.3 Flash

Oct 22, 2026 · by @Italianclownz

Context Quant Engine Decode Prefill
100K — rocmfpx recommended test build for GLM — 400.0 tok/s

I uploaded the custom GLM 5.3 Flash to hugging face and if you use the recommended test build of ROCmFPX for GLM you should get prefill above 400 tok/s over 100k tokens. Everything is explained in the model card. Should work well on Strix Halo 395+ 128 GB. Hope it works well for everyone like it does for me. If you use any of the GLM code boosts in your projects please give credit to ROCmFPX https://t.co/6HyuQbnKV2

Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, quantization, hardware state, measurement method).

Unspecified platform × deepseek v4.1-flash-1017

Oct 22, 2026 · by @xuanwo

Context Quant Engine Decode Prefill
— — — 400.0 tok/s —

迄今为止，我用过的最爽的推理服务是 @aiandcloud 部署的 deepseek，背后是最新的 v4.1-flash-1017 400 tok/s + 99% 的缓存命中，还给我送每月 $10000 额度，根本用不完（每月限额对他来说只是个数字 🤣） https://t.co/5ZbyiqgrXu

Digest-only: hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).

DGX Spark

Oct 22, 2026 · by @ai_xiaomu · third-party report

Context Quant Engine Decode Prefill
— — — 300.0 tok/s —
— — — 40.0 tok/s —

习惯了 deepseek 在线高达 300 tok/s 的流速，我已经回不去本地版 40 tok/s 的流速了。 所以我的 3 台 dgx spark 现在被我改造成了 comfyui 工厂。。。 https://t.co/fYmykP8XR3

Digest-only: confidence 0.35 below 0.60; model.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).

Unspecified platform × GPT-6.1 Sol

Oct 22, 2026 · by @LinearUncle

Context Quant Engine Decode Prefill
— — — 80.0 tok/s —

我靠，在deepseek harness(DSH)里使用magpie里codex 订阅反代出来的GPT-6.1 Sol 模型有80 tok/s 的速度，啥时候变这么快了？ https://t.co/BFDhRbJKh9

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 1017

Oct 22, 2026 · by @xiax0603 · third-party report

Context Quant Engine Decode Prefill
— — — 1521.0 tok/s —

🚨碾压Claude Opus 5.5！DeepSeek V4.1 Flash 1017跑出1521 tok/s，极速生成可玩俄罗斯方块！ Fireworks Ultrafast平台上演直观对照测试，DeepSeek V4.1 Flash 1017吞吐突破1521 tok/s，仅5.21秒输出完整可交互俄罗斯方块代码。 同一项任务，Claude Opus 5.5耗时超73秒，二者交付速度差距悬殊，足足有14倍差距啊！ 过去开发这类小游戏往往需要漫长等待，如今模型输出几乎做到随想随出。 超高吞吐不再只是纸面跑分，它会直接重塑编码智能体的实时交互体验，改写开发者的工作流。 https://t.co/dT8d7H4aRM

Digest-only: confidence 0.55 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).

Google TPU × Muse Spark Open

Oct 22, 2026 · by @GenAI_is_real

Context Quant Engine Decode Prefill
8K — sglang 770.0 tok/s —

Meta is running SGLang in production on Google's TPUs. And the stack is open source This is one of the things I love most about working in open-source AI infrastructure. SGLang-TorchTPU is the result of a remarkable collaboration between engineers at @Google, Meta's @PyTorch team, and our SGLang team at @radixark. What's especially exciting is that we're bringing much more than basic model execution to TPU. Prefix…

Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state).

RTX 3090 × DeepSeek V4.1 Flash 1017

Oct 22, 2026 · by @superalesha

Context Quant Engine Decode Prefill
— — vllm 574.0 tok/s —

3 days after this post. It keeps happening. My 3090s: 1 Linux kernel module unlocked full-speed BF16 tensor math, 82 to 155 TFLOPS on the tensor pipe without touching the driver or the 300W limit. Real prefill in Strata got +10 to 12%, the rest of the engine is the next bottleneck. DeepSeek V4.1 Flash 1017 at home: @fraserpricee runs it at 574 tok/s, 2x faster than DeepSeek's own API. vLLM matched DeepInfer's decode…

Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, quantization, context length, measurement method).

M4 Pro MacBook Pro × Flint

Oct 22, 2026 · by @QwenAIFans · third-party report

Context Quant Engine Decode Prefill
— 3-bit GGUF — 125.0 tok/s —

⚡️ You don't need 80GB of memory to run a 40B coding model locally anymore. A 24GB Mac does it. Recursive's Flint (Apache-2.0): 39.8B params, only ~2.5B active per token. Unsloth's 3-bit GGUF is 18.5GB, and they quote 125+ tok/s on an M4 Pro MacBook Pro. vs Qwen3.6-35B-A3B: 🔹 Terminal-Bench 2.1: 54.9 vs 44.9 🔹 SWE-bench Pro v2: 89.7 vs 73.8 Plot twist: dense Qwen3.8-27B still beats it (73.0 / 94.9). Catch: these are…

Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, context length, hardware state, measurement method).

MacBook Pro × Recursive Flint

Oct 22, 2026 · by @superalesha · third-party report

Context Quant Engine Decode Prefill
— — — 125.0 tok/s —
— — — 150.0 tok/s —

A new US lab just shipped a real open model. Recursive Flint. 40B coding MoE, 2.5B active, trained from scratch in under 4 months. Apache 2.0, weights on Hugging Face from day 1. 125+ tok/s on a MacBook Pro. 150+ on RTX Spark. About 7 GiB of KV cache at 262K context. They quantized the lm_head to 4 bits, the part everyone keeps in high precision, and it stays within 0.7 points of BF16. An automated research system…

Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, context length, hardware state, measurement method).

Fireworks Ultrafast × DeepSeek V4.1 Flash 1017

Oct 22, 2026 · by @NFT_Chen · third-party report

Context Quant Engine Decode Prefill
— — fireworks ultrafast 1521.0 tok/s —

😱逆天了！DeepSeek V4.1 Flash 1017 在Fireworks Ultrafast上跑出1521+ tok/s！ 11秒就能写出完整可玩的Tetris游戏，比Claude Opus 5.5快7倍。 以前写个小游戏还要等半分钟，现在基本是“想”完就出。 开源模型推理速度已经来到这种程度了， coding agent 的交互体验要被彻底改写。 #DeepSeek #AI #推理速度 #Fireworks #开源大模型 #Tetris #梁文峰 https://t.co/pXfSNWIj4L https://t.co/Tlunbc44ab

Digest-only: confidence 0.00 below 0.60; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).

Unspecified platform × DeepSeek V4.1-Flash

Oct 22, 2026 · by @hsu_steve · third-party report

Context Quant Engine Decode Prefill
700K FP8 — 230.0 tok/s —

China produces more AI tokens. The US produces more AI value per token. Ascend stack significant in 2027. [ Research using Claude ] Two numbers get mixed up in token debates. Total tokens count everything a model processes, mostly input: prompts, documents and long agent contexts it reads. Output tokens are the text it actually generates, a small fraction of the total and the part that costs the most compute. Total…

Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, hardware state, measurement method).

M3 Ultra × DeepSeek-V4.1-Flash-1017

Oct 22, 2026 · by @rapidmlx

Rapid-MLX 0.16.7 is out: DeepSeek V4.1 Flash runs up to ~2x faster on a Mac. Our quant of DeepSeek-V4.1-Flash-1017 (GPTQ v2) now decodes at 33–45 tok/s single-stream on an M3 Ultra. Same weights, faster engine: lighter drafting, fused mixed-bit MoE, and a vocab head read once per verify round. Every token is still verified by the full model. pip install -U rapid-mlx rapid-mlx serve deepseek-v41-flash-1017 Try it…

Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (context length, hardware state, measurement method).

Also in this window

The 12 most recent posts are shown in full above; 25 posts from the same window are listed here in short form, each with its source post.

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

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.

↑