X local-model bench digest — Oct 28, 2026
As of Oct 28, 2026 — 47 posts collected from X. 98 measured runs catalogued, 47 kept to the digest only. Latest source post Oct 28, 2026 10:15 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 x2 × DeepSeek-V4.1-Flash
Oct 28, 2026 · by @Blackwellboy
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 113.0 tok/s | — |
Week one of BLACKFOLD: several hundred downloads and so much love from you lot 🏎️ It's been thrown at coding, pentest ranges and voxel games, and one tester reckons it may be his new daily driver. Uncensored DeepSeek-V4.1-Flash. One of the strongest TP2 builds on two DGX Sparks. 113 tok/s code Tool calls clean to 250K 7 fixes merged upstream Free weights + recipe. Have a crack 👇
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 3090 x4 × DeepSeek V4.1 Flash
Oct 28, 2026 · by @superalesha
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 57.0 tok/s | — |
DeepSeek doesnt need a subscription. The weights are MIT. Every other lab should be embarrassed. You pay per token on their API, or you download V4.1 Flash and never pay per token again. I run it on my 4x 3090 at 57 tok/s and nobody can cap me. Now look at the rest of the market this month. SemiAnalysis caught Anthropic A/B testing usage limits, and some accounts got about 20% less monthly value without a word. My…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeekV4.1-Flash-1017
Oct 28, 2026 · by @WojakTheDev
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 4096 | 2.00 bpw | strata 0.2.5 | 5.0 tok/s | 110.0 tok/s |
| 4096 | 2.00 bpw | strata 0.2.5 | 10.0 tok/s | 155.0 tok/s |
| 8192 | 2.00 bpw | strata 0.2.5 | 23.0 tok/s | 340.0 tok/s |
| 8192 | 2.00 bpw | strata 0.2.5 | 75.5 tok/s | 425.0 tok/s |
| 4096 | 2.00 bpw | strata 0.2.7 | 20.0 tok/s | 87.0 tok/s |
| 8192 | 2.00 bpw | strata 0.2.7 | 20.0 tok/s | 125.0 tok/s |
run DeepSeek V4.1 Flash local on ONE GPU Strata · disk-offloading · ~46–64GB RAM class boxes how: 1. grab Strata (https://t.co/TW1QSrhgJT) 2. one-click / strata launch --family deepseek41 3. OpenAI-compatible API on localhost works with Claude Code · Codex · OpenCode needs lots of free SSD space · expect a few tok/s on consumer GPUs
Digest-only: hardware.name missing; incomplete methodology (hardware state, measurement method).
Unspecified platform × GLM 5.3 Flash
Oct 28, 2026 · by @runinfrai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 1M | — | — | 254.0 tok/s | — |
glm 5.3 flash just got 51% cheaper on runinfra $0.05 per 1M input, down from $0.11. $0.01 per 1M cached input, down from $0.03. $0.22 per 1M output, down from $0.45 99.3% cache hit rate over the last 24h. most of your input tokens are billing at the cached rate that puts a full 1M context prompt at around 1 cent on a cache hit same speed: 254 tok/s, 703ms TTFT discount runs until nov 4, then standard rates resume…
Digest-only: confidence 0.00 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 3090 × Qwen3.8-Flash-Next
Oct 28, 2026 · by @NFT_Chen
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | IQ3_XXS | strata 0.2.3 | 102.0 tok/s | — |
| — | — | strata 0.2.3 | 55.0 tok/s | — |
⚡️秒懂！本地跑 Qwen 3.8 Flash Next 性能远超Qwen官方API，一张 3090 速度从哪来？配置怎么配？ 本地跑 Qwen3.8-Flash-Next（Strata 专家卸载 + 4bit 动态量化），一张二手 3090 就能接近官方 API 2 倍速度，核心就这几件事： 速度怎么来的？ 🔹MoE + 专家卸载：Strata 把热门专家放显存，其余放内存，命中率 93-96%，大部分 token 不用全量跑。 🔹MTP 投机解码：一次猜 4 个 token 再验证，接受率高，直接加速。 🔹框架优化：Strata 0.2.3 针对 3090（sm_86）做了专门优化，再快约 9%。 🔹内存充足是前提：专家没被挤到 SSD，带宽才不拖后腿。 关键配置清单 1️⃣显卡：RTX 3090 24GB ×1（显存实际占用约 22.5GB） 2️⃣内存：≥96GB（推荐 128GB），实际占用 62-70GB。64GB…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (context length, measurement method).
single GPU (unspecified) × Clover-35B-A3B-Preview
Oct 28, 2026 · by @NFT_Chen · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | BF16 | — | 680.0 tok/s | — |
🚨最新！中国团队扩散智能（DiffuSpace）扔出重磅模型Clover-35B-A3B-Preview 首个开源生产级扩散语言模型（dLLM），35B总参数，激活仅3B，单卡BF16直接飙到680 tokens/s！ 关键亮点： 🔹 生成范式颠覆：不是Qwen/GPT那种逐token生成，而是并行扩散+实时修正，Agent多轮延迟直接腰斩 🔹 Agent性能炸裂：SWE-bench Verified 73.6（略超同尺寸Qwen3.6-35B-A3B的73.4）、Terminal-Bench 51.7，六项Agent评测全面领先 🔹 速度提升：同规模自回归模型5-10倍提速，尤其适合代码Agent、网页自动化、实时生成UI 🔹 完全开源：HF + ModelScope + 官网同步上线，支持SGLang/llama.cpp，Apache-2.0随便玩 从Dream…
Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, context length, hardware state, measurement method).
DGX Spark × Qwen3.8-Flash-Next
Oct 28, 2026 · by @atasrb
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 66.0 tok/s | — |
| — | — | — | 96.0 tok/s | — |
| — | — | — | 11.0 tok/s | — |
| — | — | — | 103.0 tok/s | — |
| — | 2.15 bpw | — | 28.0 tok/s | — |
| — | 3.06 bpw | — | 21.0 tok/s | — |
| — | — | — | 99.0 tok/s | — |
What fits on one DGX Spark, and what needs two. Only models I ran myself, one user, real prompts. One Spark: Qwen3.8-Flash-Next at 66 tok/s on JSON. Nemotron 3.5 Bolt at 96 with its draft model. A dense 27B at 11. Two Sparks: DeepSeek-V4.1-Flash (763B) at 103 tok/s, but only 7-9 GiB stays free on each Spark. MiniMax-M3.1-Flash (428B) at 27-29 with 2.15 bpw, 20-22 with 3.06 bpw. Qwen3.8-Flash-Next goes from 66 to 99.
Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, context length, measurement method).
RTX 3090 x4 × DeepSeek V4.1 Flash
Oct 28, 2026 · by @superalesha
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 57.0 tok/s | — |
OpenAI launched GPT-6.1 Terra at $1 in, $5 out. Half the price of Sol. Say thank you to Hangzhou. Nobody at OpenAI woke up generous. In October DeepSeek V4.1 Flash went into Cursor at $0.60 to $0.98 per task, Kimi K3.1 Pro launched at half the price of Opus 5.5, and suddenly OpenAI and Anthropic both cut prices in the same month. My view: cheap Chinese weights forced it. You dont even need to touch a Chinese model…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
M5 MacBook Pro × Qwen3.5 9B
Oct 28, 2026 · by @NFT_Chen · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | 4-bit | uzu | 92.1 tok/s | — |
| — | 4-bit | llama.cpp | 22.0 tok/s | — |
| — | 4-bit | mlx | 25.1 tok/s | — |
💥卧槽！本地LLM推理迎来真正突破！uzu引擎比llama.cpp快4倍，M5 Mac上9B模型直接飙到92 tok/s！ 一次解码平均产出7.5个token，彻底撕开本地推理的内存带宽天花板！ 核心亮点： 🔹实测碾压：在基础版M5 MacBook Pro上跑Qwen3.5 9B（4-bit），llama.cpp仅22.0 tok/s、MLX 25.1 tok/s，uzu直接92.1 tok/s，约4倍速度。 🔹技术黑科技：采用激进投机解码（Speculative Decoding）。小模型DFlash一次性提案未来16个位置，56.7M参数的Weaver适配器构建连贯树，主模型单次pass验证整棵树并保留最长正确路径。普通解码每pass仅1 token，uzu平均7.5 token。 🔹硬件协同设计：Rust + Metal内核，专为Apple Silicon优化，充分利用M5的int8矩阵路径和统一内存。支持Python…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
RTX 3090 x2 × GLM-5.3-Flash
Oct 28, 2026 · by @TranHuuTaiAu
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 262K | — | — | 41.5 tok/s | — |
Picked up 2x MSI RTX 3090 Ventus and turned them into a home rig for GLM-5.3-Flash (321B MoE). 38-45 tok/s decode, ~1s to first token, 262K context. Core Ultra 5 + 128GB DDR5, nothing fancy. Followed @catonooka's maya-s24-turbo recipe. Full specs in the last image. https://t.co/Axlw5JkTz1
Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, quantization, hardware state, measurement method).
Unspecified platform × Qwen
Oct 28, 2026 · by @hasantoxr · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | vllm | 700.0 tok/s | — |
A 35B open model just hit ~700 tokens per second That's 5-10x faster than a regular model the same size It's called Clover 35B-A3B from DiffuSpace founded by @ikekong and it doesn't write like any LLM you've used Every model you use today (ChatGPT, Claude, Qwen) writes one token at a time. Left to right. Each one waiting on the last Clover is a diffusion model. It drafts a whole block of text at once, then refines…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
M4 Max × LFM3-1.7B
Oct 28, 2026 · by @tomgreenwald
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | magnitude | 280.0 tok/s | — |
𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 running locally at 280 tok/s on my M4 Max Generation is real time (only my typing is sped up) Surprisingly coherent for its size, and less prone to overthinking than the small Qwen models Day 1 support in Magnitude, shipped today 😁 Try it now: https://t.co/Agab8hUc82
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 35 posts from the same window are listed here in short form, each with its source post.
- DGX Spark x4 × DeepSeek-V4.1-Flash — 7324.0 tok/s prefill (2 measurements) · @tenhkspark
- RTX PRO 6000 × LFM3 — 29212.0 tok/s decode (3 measurements) · @Nateemerson
- Unspecified platform × DeepSeek V4.1 Flash — 1500.0 tok/s decode (1 measurement) · @brucexu_eth
- RTX 3060 × Qwen3.8-35B-A3B — 50.0 tok/s decode (1 measurement) · @Oluwaphilemon1
- RTX 3090 × Qwen3.8-27B — 1035.0 tok/s decode (2 measurements) · @QwenAIFans
- RTX PRO 6000 x2 × DeepSeek-V4-Flash — 1169.0 tok/s decode (1 measurement) · @autonomous_labs
- DGX Spark × Ling-3.0-flash — 53.9 tok/s decode (2 measurements) · @NeoAIForecast
- RTX 3090 × LFM3-1.7B — 384.0 tok/s decode (2 measurements) · @Layton_Gott
- RTX 5060 Ti x2 × Qwen Next Flash — 65.0 tok/s decode (1 measurement) · @RRunner144
- DGX Spark x4 × DeepSeek V4.1 Flash 1017 — 110.9 tok/s decode (6 measurements) · @wei_wang
- M5 MacBook Pro × Qwen3.5 9B — 92.1 tok/s decode (3 measurements) · @MLBear2
- RTX 3090 x4 × MiniMax-M3.1-Flash — 23.0 tok/s decode (3 measurements) · @superalesha
- RTX 5090 × Ornith 1.5 122B — 303.0 tok/s decode (2 measurements) · @witcheer
- M4 Max × Qwen3.6 35B-A3B Uncensored — 115.0 tok/s decode (1 measurement) · @tomgreenwald
- RTX 3060 × LFM3-1.7B — 244.0 tok/s decode (3 measurements) · @DogukanUrker
- TITAN RTX × Qwen3.8-27B — 7752.0 tok/s prefill (3 measurements) · @QwenAIFans
- Arm SME2 CPU (2 threads) × LFM3-1.7B — 50.0 tok/s decode (2 measurements) · @liquid_richie
- M5 Pro 64GB × LFM3-1.7B — 184.0 tok/s decode (5 measurements) · @pratikg
- DGX Spark x4 × DeepSeek v4.1 Flash 1017 — 107.0 tok/s decode (1 measurement) · @norsyx
- Arm CPU (SME2) × LFM3-1.7B — 50.0 tok/s decode (2 measurements) · @Arm
- Mac Studio 256GB × 5.3 flash — 85.0 tok/s decode (1 measurement) · @rez0__
- RTX 6000 x4 × DeepSeek v4.1 1017 — 340.0 tok/s decode (2 measurements) · @ZenMagnets
- DGX Spark × Unlimited-OCR-2 — 659.0 tok/s decode (6 measurements) · @ivanfioravanti
- Unspecified platform — 850.0 tok/s decode (1 measurement) · @SemiAnalysis_
- Ryzen AI Max (Strix Halo) × DeepSeek-V4.1-Flash-1017 — 300.0 tok/s prefill (1 measurement) · @NFT_Chen
- RTX PRO 6000 × OXR-1.0 — 558.0 tok/s decode (1 measurement) · @rohanpaul_ai
- P100 × Qwen 3.8 27b — 26.8 tok/s decode (1 measurement) · @iamMrDuncan
- AMD Instinct MI355X x8 × DeepSeek V4.1 Flash — 2100.0 tok/s decode (1 measurement) · @lithos_ai
- RTX 4090 Laptop 16GB × Qwen3.8-Flash-Next (Gyro-S, abliterated) — 92.5 tok/s decode (1 measurement) · @mashzyna
- M5 MacBook × Qwen3.5 9B — 92.1 tok/s decode (3 measurements) · @ashxhart
- M5 Pro 64GB × Nemotron 3.5 Bolt E4B — 63.0 tok/s decode (1 measurement) · @pratikg
- Radeon AI PRO R9700 x4 × DeepSeek-V4.1-Flash — 1350.0 tok/s prefill (4 measurements) · @uctive
- Mac Studio 128GB × DeepSeek 1017 — 76.0 tok/s decode (1 measurement) · @ohayou792
- Unspecified platform × DeepSeek V4.1 Flash 1017 — 750.0 tok/s decode (1 measurement) · @BoundlessHQ
- RTX 3090 x4 × DeepSeek V4.1 Flash — 57.0 tok/s decode (1 measurement) · @superalesha
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
- @Blackwellboy — Oct 28, 2026
- @superalesha — Oct 28, 2026
- @WojakTheDev — Oct 28, 2026
- @runinfrai — Oct 28, 2026
- @NFT_Chen — Oct 28, 2026
- @NFT_Chen — Oct 28, 2026
- @atasrb — Oct 28, 2026
- @superalesha — Oct 28, 2026
- @NFT_Chen — Oct 28, 2026
- @TranHuuTaiAu — Oct 28, 2026
- @hasantoxr — Oct 28, 2026
- @tomgreenwald — Oct 28, 2026
- @tenhkspark — Oct 28, 2026
- @Nateemerson — Oct 28, 2026
- @brucexu_eth — Oct 28, 2026
- @Oluwaphilemon1 — Oct 28, 2026
- @QwenAIFans — Oct 28, 2026
- @autonomous_labs — Oct 28, 2026
- @NeoAIForecast — Oct 28, 2026
- @Layton_Gott — Oct 28, 2026
- @RRunner144 — Oct 28, 2026
- @wei_wang — Oct 27, 2026
- @MLBear2 — Oct 27, 2026
- @superalesha — Oct 27, 2026
- @witcheer — Oct 27, 2026
- @tomgreenwald — Oct 27, 2026
- @DogukanUrker — Oct 27, 2026
- @QwenAIFans — Oct 27, 2026
- @liquid_richie — Oct 27, 2026
- @pratikg — Oct 27, 2026
- @norsyx — Oct 27, 2026
- @Arm — Oct 27, 2026
- @rez0__ — Oct 27, 2026
- @ZenMagnets — Oct 27, 2026
- @ivanfioravanti — Oct 27, 2026
- @SemiAnalysis_ — Oct 27, 2026
- @NFT_Chen — Oct 27, 2026
- @rohanpaul_ai — Oct 27, 2026
- @iamMrDuncan — Oct 27, 2026
- @lithos_ai — Oct 27, 2026
- @mashzyna — Oct 27, 2026
- @ashxhart — Oct 27, 2026
- @pratikg — Oct 27, 2026
- @uctive — Oct 27, 2026
- @ohayou792 — Oct 27, 2026
- @BoundlessHQ — Oct 27, 2026
- @superalesha — Oct 27, 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.