X local-model bench digest — Oct 15, 2026
As of Oct 15, 2026 — 34 posts collected from X. 84 measured runs catalogued, 1 published to the board, 33 kept to the digest only. Latest source post Oct 15, 2026 07:47 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.
MacBook Pro M5 Pro × Qwen 3.8 Flash Next
Oct 15, 2026 · by @aartiles24
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | sushi | 41.4 tok/s | — |
If you're not in a hurry, you don't need a frontier model to build with the anatomy skill. Here's a microwave built entirely on a MacBook Pro M5 Pro (64 GB) with Qwen 3.8 Flash Next, quantized to ~2.6 bits, served by Sushi and driven by Pi. One long end-to-end task with zero human input to the model: no prompts, no hints, not a line of the project's code from me. - 17 h 20 min run time - 76 agent sessions, 23/23…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
DGX Spark × Qwen3.8-27B
Oct 15, 2026 · by @0xBakeer · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | vllm | 49.9 tok/s | — |
TandemLLM now has a real manual. 260 pages 📖 It follows one request from the client to the last token. It covers the conductor that plans each turn, and how speculative decoding takes Qwen3.8-27B from 13.7 to 49.9 tok/s on a single DGX Spark with StairCut. It also covers caches that resume bit for bit after a restart, memory notes your agents keep between sessions, and one dashboard in front of vLLM, SGLang, OpenAI…
Digest-only: confidence 0.35 below 0.60; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
M5U × Qwen 27B
Oct 15, 2026 · by @ddalcu
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | mlx | — | 2280.0 tok/s |
| — | — | mlx | 462.0 tok/s | — |
Using @anemll's Forge repo as reference, MLX-Serve can soon run Qwen 27B with 36 Users on one M5U * Three prefills run parallel (2,280 tok/s combined) * 36 users in parallel (462 tok/s combined) * Yes you can have subagents now https://t.co/SxnwURwAsl
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
PS5 × Qwen 3.8 27B
Oct 15, 2026 · by @NFT_Chen
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 64K | IQ2_XXS | llama.cpp | 20.7 tok/s | — |
| — | — | llama.cpp | 61.0 tok/s | — |
| — | — | llama.cpp | 124.0 tok/s | — |
🤯简直没天理了！越狱版PS5直接跑通270亿参数的Qwen 3.8 27B，变身本地AI服务器！ PS5LM Plus 把llama.cpp做成原生应用，通过Vulkan + RADV驱动调用PS5 GPU，可用内存从原来的约6GB提升到约11.4GB，终于能塞进重度量化后的27B模型。 关键亮点 🔹Qwen 3.8 27B（IQ2_XXS约6.8GB + 64K上下文）：约20.7 tok/s 🔹小模型更离谱：Gemma 4 E4B约61 tok/s，Llama 3.2 3B高达124 tok/s 🔹GPU几乎打满，CPU几乎闲着，整机功耗约86W，温度仅54°C 🔹电视上有实时仪表盘（速度、功耗、温度、内存占用） 🔹支持从Hugging Face直接下载模型，并提供OpenAI兼容API，手机/电脑同一局域网即可调用，真正变身家庭AI服务器 快速上手步骤：…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
DGX Spark x2 × DeepSeek-V4.1
Oct 15, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 256K | — | vllm v1.1 | 141.0 tok/s | — |
| 256K | — | vllm v1.1 | 103.0 tok/s | — |
| 255K | — | vllm v1.1 | — | 2090.0 tok/s |
i couldn't leave it alone on monday i released the fastest vllm setup for deepseek v4.1 on 2x dgx sparks 141 tok/s on structured 103 on code full 256k context v1.1 made prefill 18% faster, about 2,090 tok/s, and a 255k token prompt that used to take 183 seconds before the first token now takes 160, so long sessions feel a lot better than what i first shipped the weights have 114 downloads since monday, that's the…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (quantization, hardware state, measurement method).
M5 Ultra × strata-mlx
Oct 15, 2026 · by @yibie
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 256 | Q2_0 | mlx | 110.0 tok/s | — |
| 256 | Q2_0 | mlx | 170.0 tok/s | — |
| 256 | Q2_0 | mlx | 156.0 tok/s | — |
| 256 | IQ3_S | mlx | 112.0 tok/s | — |
| 256 | IQ3_S | mlx | 161.0 tok/s | — |
| 256 | IQ3_S | mlx | 137.0 tok/s | — |
strata-mlx 更新：合并了社区贡献者 vendyluo 的两个 PR，写回答更快了。 M5 Ultra 上带草稿头写回答（作者实测，256 token，三轮取中位数）： ・Q2_0：中文 86 → 110 tok/s，代码 127 → 170，英文散文 127 → 156 ・IQ3_S：中文 95 → 112 tok/s，代码 127 → 161，英文散文 118 → 137 PR 9：把一次前向里十组互相等待的小内核合并，输出与原来逐位一致。M4 Max 上每次前向 1 个 token 22.5 → 20.8 ms，4 个 token 35.0 → 31.6 ms。 PR 10：启动后在后台把 28.8 GB 的 n-gram 表读进缓存，写回答时不再等缺页。 两台机器交叉验证时还发现，同一个参数在 M5 和 M4 上的最佳值正好相反，接下来会按芯片自动选择。 感谢 vendyluo，这是他提交的第三个 PR。…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state).
RTX 3060 × Qwen3.8-35B-A3B
Oct 15, 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 |
Qwen3.8-35B-A3B is reportedly running at around 50 tokens/sec on an RTX 3060 with just 12GB VRAM and 16GB system RAM. The reported setup also delivers ~550 tokens/sec in prompt processing, using a Q4_K_M quant and a configured 262K context window. Hardware and configuration: GPU: RTX 3060 12GB System RAM: 16GB Quantization: Q4_K_M Generation: ~50 tok/s Prompt processing: ~550 tok/s Context: 262K tokens Runtime…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
PS5 × Qwen3.8-27B
Oct 15, 2026 · by @jurlycat
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | llama.cpp | 20.7 tok/s | — |
A jailbroken PS5 is running a 27B LLM at ~20 tokens/sec. 💀 PS5LM Plus runs Qwen3.8-27B locally on the console’s GPU using llama.cpp + Vulkan. The trick: turning the original CPU payload into a native PS5 app, unlocking GPU access and ~11.4 GiB of usable memory. The developer reports 20.7 tok/s with a ~2-bit quant. It also serves an OpenAI-compatible API, so other devices on the network can use the PS5 as a local AI…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, measurement method).
DGX Spark x2 × DeepSeek-V4.1-Flash
Oct 15, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 74.0 tok/s | — |
| — | — | — | 141.0 tok/s | — |
| — | — | — | 103.0 tok/s | — |
i've been running deepseek v4.1 flash on my 2x dgx sparks for about a week now and oh my dear lord this model is so so so much closer than you think don't judge it by its frontend capability, put it on backend and real work and watch how it drives, there are sessions where i feel like i'm talking to opus 5.5 and i completely forget that frontier exists, that's the part that matters for how i work, i pair engineer…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
AMD Radeon 8065S × GLM-5.3-Flash-Maya-S
Oct 14, 2026 · by @PeasantSmith
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 32768 | — | project maya engine v1.1.10 | 29.6 tok/s | — |
| 32768 | — | project maya engine v1.1.10 | 23.2 tok/s | — |
| 8192 | — | project maya engine v1.1.10 | — | 278.0 tok/s |
Project Maya v1.1.10: Strix/Gorgon Halo release • Answers +50%: up to 29.6 tok/s (was 19.7) • After 8K prompts +26%: 23.2 tok/s (was 18.4) • PromptP +10%: 278 tok/s at 8K (was 253) Same answers, only faster. Measured on a Gorgon Halo, 192 GB. https://github.com/mw00/project-maya/releases
Full methodology stated.
Mac Mini + RTX 3090 × Qwen3.8-Next-Flash
Oct 14, 2026 · by @0xSero
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | EXL3-3bits | — | 31.0 tok/s | — |
This is running on a 16gb mac mini and 3090 Incredible model beats most of the mid stuff, qwen3.8-next-flash in exl3-3bits over half the model is being streamed from NVME it's 31 tok/s that is faster than Codex as of late.. for 2600$ I feel like my mission needs updating https://t.co/qVxtso2EIE
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
M4 Pro 24GB + iPhone 17 Pro Max × Qwen3.8-27B
Oct 14, 2026 · by @BrianRoemmele
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 16K | IQ4_XS | llama.cpp | — | 157.0 tok/s |
| 32K | IQ4_XS | llama.cpp | — | 130.0 tok/s |
| 48K | IQ4_XS | llama.cpp | — | 113.0 tok/s |
iPhone 17 As A MacBook Pro Accelerator! Apple placed a neural accelerator in each GPU core of the A19 Pro and stated that the result is MacBook Pro levels of compute in an iPhone, suited to GPU-intensive AI workloads. The claim appears in the September 2025 keynote for the iPhone 17 series. Independent measurements now show what that compute does when the phone is attached to a larger machine. A developer running a…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version).
Also in this window
The 12 most recent posts are shown in full above; 22 posts from the same window are listed here in short form, each with its source post.
- DGX Spark x2 × GLM-5.3-Flash — 100.4 tok/s decode (4 measurements) · @MehdiHacks
- RTX 5090 × LFM2.5-24B-A2B — 351.5 tok/s decode (2 measurements) · @niklaslenz_ai
- RTX 3090 — 27.5 tok/s decode (1 measurement) · @needmorevram
- DGX Spark x2 × DeepSeek-V4.1-Flash — 141.0 tok/s decode (9 measurements) · @yume_arasaki
- Unspecified platform × gemma 4 — 0.4 tok/s decode (1 measurement) · @SadCreatorTalks
- RX9060XT x2 × Qwen3.8-Flash-Next — 58.9 tok/s decode (4 measurements) · @tamanekokoro
- PS5 × Qwen 3.8 27B — 124.0 tok/s decode (3 measurements) · @Pirat_Nation
- RTX 3060 Ti × llama.cpp — 130.0 tok/s decode (2 measurements) · @rna4219
- MI355X x8 × DeepSeek V4 Pro 0813 — 647.0 tok/s decode (7 measurements) · @ArtificialAnlys
- PS5 (jailbroken) × Qwen 3.8 27B — 61.0 tok/s decode (2 measurements) · @AVAmirVA
- RTX PRO 6000 × GLM-5.3-Flash — 69.9 tok/s decode (3 measurements) · @chenran818
- M5 Max × Qwen3.8-27B — 882.0 tok/s prefill (1 measurement) · @trymirai
- DGX Spark × GLM-5.3 Flash M288 Mosaic 12L — 30.7 tok/s decode (2 measurements) · @filicroval
- DGX Spark × LFM2.5-24B-A2B — 151.0 tok/s decode (1 measurement) · @0xBakeer
- RTX 5090 Mobile × MiniCPM-V 4.7 — 272.0 tok/s decode (3 measurements) · @sudoingX
- Unspecified platform × Qwen-3 — 0.4 tok/s decode (1 measurement) · @DavidSacks
- RTX 3090 × MoE — 273.0 tok/s decode (2 measurements) · @coffeecup2020
- Unspecified platform × LFM2.5-24B-A2B — 184.0 tok/s decode (2 measurements) · @wei_wang
- Unspecified platform × Qwen — 20.0 tok/s decode (1 measurement) · @kimmonismus
- Unspecified platform — 470.0 tok/s decode (1 measurement) · @firesidealpha
- Unspecified platform × Qwen — 20.0 tok/s decode (1 measurement) · @pupposandro
- Unspecified platform × Qwen — 124.0 tok/s decode (2 measurements) · @mylifcc
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
- @aartiles24 — Oct 15, 2026
- @0xBakeer — Oct 15, 2026
- @ddalcu — Oct 15, 2026
- @NFT_Chen — Oct 15, 2026
- @sudoingX — Oct 15, 2026
- @yibie — Oct 15, 2026
- @Oluwaphilemon1 — Oct 15, 2026
- @jurlycat — Oct 15, 2026
- @sudoingX — Oct 15, 2026
- @PeasantSmith — Oct 14, 2026
- @0xSero — Oct 14, 2026
- @BrianRoemmele — Oct 14, 2026
- @MehdiHacks — Oct 14, 2026
- @niklaslenz_ai — Oct 14, 2026
- @needmorevram — Oct 14, 2026
- @yume_arasaki — Oct 14, 2026
- @SadCreatorTalks — Oct 14, 2026
- @tamanekokoro — Oct 14, 2026
- @Pirat_Nation — Oct 14, 2026
- @rna4219 — Oct 14, 2026
- @ArtificialAnlys — Oct 14, 2026
- @AVAmirVA — Oct 14, 2026
- @chenran818 — Oct 14, 2026
- @trymirai — Oct 14, 2026
- @filicroval — Oct 14, 2026
- @0xBakeer — Oct 14, 2026
- @sudoingX — Oct 14, 2026
- @DavidSacks — Oct 14, 2026
- @coffeecup2020 — Oct 14, 2026
- @wei_wang — Oct 14, 2026
- @kimmonismus — Oct 14, 2026
- @firesidealpha — Oct 14, 2026
- @pupposandro — Oct 14, 2026
- @mylifcc — Oct 14, 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.