X local-model bench digest — Oct 13, 2026
As of Oct 13, 2026 — 33 posts collected from X. 89 measured runs catalogued, 2 published to the board, 31 kept to the digest only. Latest source post Oct 13, 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 × GLM-5.3
Oct 13, 2026 · by @0xSero · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 100.0 tok/s | — |
New machine new me. This one is about 12,000$ and can run GLM-5.3-Flash and some 100 tok/s https://t.co/YJX0hZXfac
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark x2 × DeepSeek-V4.1-Flash
Oct 13, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 8K | — | custom dgx spark backend (routed v1.1 | — | 2090.0 tok/s |
| 32K | — | custom dgx spark backend (routed v1.1 | — | 2092.0 tok/s |
| 32K | — | custom dgx spark backend (routed v1.1 | 103.0 tok/s | — |
| 32K | — | custom dgx spark backend (routed v1.1 | 57.9 tok/s | — |
| 32K | — | custom dgx spark backend (routed v1.1 | 141.0 tok/s | — |
update! deepseek v4.1 flash just got faster at reading v1.1 of my 2x dgx spark recipe is out and the prompt side is up to 18% quicker 🐋 8k prompt, 1,763 tok/s → 2,090 tok/s prefill now 32k prompt, 1,773 tok/s → 2,092 tok/s prefill now decode unchanged, 103.0 / 57.9 / 141.0 tok/s on code / prose / structured, dspark on an agent hands it a 32k context and the first token lands in 15.7 s now, that's the part you feel…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (quantization, hardware state, measurement method).
Unspecified platform
Oct 13, 2026 · by @yuhasbeentaken · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 50.0 tok/s | — |
let me add the main downsides of opencode go compared to deepseek’s direct api. a few things i’ve noticed: 1. the speed gap is real. direct deepseek is often noticeably faster, sometimes by around 50 tok/s. 2. i don’t think quantization is the only possible reason. opencode may simply be sharing more users across the same inference capacity. 3. that would explain why performance can change depending on the time of…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; model.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × GLM-5.3
Oct 13, 2026 · by @tinygrad · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 250.0 tok/s | — |
Sorry, I got the price wrong for the GLM-5.3 flash machine. 250 tok/s should be doable with decent software. Buy today! https://t.co/OhC0H2RLdA
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark
Oct 13, 2026 · by @_nodelay · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 50.0 tok/s | — |
| — | — | — | 11.0 tok/s | — |
2x dgx spark에서 돌아가는 GLM 5.3 flash 의 속도가 50 tok/s ... GPT 6 Astra 11 tok/s ... 이거 뭐 일을 두번하더라도 로컬을 쓰겠는데요. openAI 뭐하는거냐 빨리 이름 openSI 로 바꾸고 속도 개선이나 해라!
Digest-only: confidence 0.35 below 0.60; model.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
gaming laptop × Kimi K3
Oct 13, 2026 · by @superalesha
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 0.1 tok/s | — |
9.4 seconds per token. Thats the real speed of the "2.8 TRILLION parameter model on a gaming laptop". Its in the engine's own table. A 300 token answer is 47 minutes. And this guy says he cancelled $700 of subscriptions the same night. Bullshit. Colibri is real and its a beautiful piece of engineering. Its author is honest about what it is. Read his own table: Kimi K3: 2.8T parameters, 104B active per token…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
Unspecified platform
Oct 13, 2026 · by @aigclink · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 17.4 tok/s | — |
今天上新的模型Meta的Muse Wonder、智谱的GLM-Legal中，面壁开源的MiniCPM-V 4.7值得一看，端侧多模态的部署经济学被改写了 MiniCPM-V 4.7 35B-A3B用量化压缩+LES调度，实现了手机跑35B MoE，推理速度达到每秒十几个Token 很多端侧加速方案是砍精度换速度，代价是掉点，LES在相对基线精度不动的前提下，靠调度把IO等待藏起来，最终实现从2.97达到17.39 tok/s（5.86x） iPad Pro M4（16GB）上达到17+Token/s；iPhone 17 Pro Max上达到13+Token/s 视觉侧还支持4倍、16倍混合Token压缩，进一步降低多模态处理开销 这进一步利好隐私敏感场景 #端侧模型 #MiniCPMV47
Digest-only: confidence 0.35 below 0.60; hardware.name missing; model.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark x2 × Qwen3.8-Flash-Next
Oct 13, 2026 · by @jmacftw
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | NVFP4 | atlasinference | — | 4043.0 tok/s |
If you want to run one of the best versions of Qwen 3.8 Flash Next on 2 DGX Sparks with @NVIDIAAI NVFP4 weights, blazing fast 4043 tok/s prefill and awesome decode and concurrency on a fully native inference engine (@AtlasInference), this is an excellent option! It’s a great local workhorse and what we often run for continuous entity cognition. https://t.co/ygQT3PzpIH
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
Unspecified platform × Qwen
Oct 13, 2026 · by @ZinkLu2 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 60.0 tok/s | — |
Claude 被封了，最近把其他模型又用了一轮，说下主观感受： GPT 不说了，全系模型都感觉缩水了，无论是智商还是审美都是一降再降，而且速度这么慢，重置不重置的意义不大，反正用不完。 GLM 5.3 是目前写代码最强的，官网 API 推理速度能到 45~60 tok/s，干的活真的又快又好，除了有点贵基本没啥毛病。 K3 思考时间很长，干活总差个临门一脚，性价比不高了。 Qwen 3.8 Max 挺不错的，但不突出，我总结不出啥优势，而且套餐不经用。（但期待 Qwen4，我目前 Qwen3.8 27B 本地部署的版本真的满足我日常使用了） MiniMax M3.1-Flash 是无人在意的角落——估计偷偷蒸馏了不少数据，有种 opus 的即视感；不过之前被 M3 伤过，目前还没敢大量用。 DeepSeek Flash 算听话，自动化场景或者办公场景，你陪他慢慢打磨，能稳定跑一些任务，但说真的，有人在用它写代码吗？…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 2070 Super × unknown 27b model
Oct 13, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 64K | — | — | 52.6 tok/s | — |
| 131K | — | — | 50.1 tok/s | — |
| 262K | — | — | 67.3 tok/s | — |
if you own a gpu with 8gb to 16gb of vram, this repo is for you i'm sudo, i bench local ai on everything from 8gb cards to 128gb sparks to a 1tb nodes and publish the numbers from the ground to the top, this one is for the ground floor one line installs a 27b local ai on your gaming gpu, linux or windows, nvidia or amd. it finds your card, picks the right build and context, checks every file and starts it, then you…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state, measurement method).
Unspecified platform
Oct 13, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 52.6 tok/s | — |
| — | — | — | 50.1 tok/s | — |
| — | — | — | 67.3 tok/s | — |
local ai for everyone! nvidia or amd, linux or windows, any card with 8gb or more, just one line installs it all. rtx or rx, both get a crown here stop renting your thinking every prompt you type into a cloud chatbot lands on someone else's server, under their rules, at their price, and they can change both whenever they want. the model you think with can live on your own vram, on the gaming gpu you might already…
Digest-only: hardware.name missing; model.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark x2 × DeepSeek-V4.1-Flash
Oct 13, 2026 · by @sudoingX
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 262K | EXL3 2.70bpw (per-expert variabl | vllm | 141.3 tok/s | — |
| 262K | EXL3 2.70bpw (per-expert variabl | vllm | 103.0 tok/s | — |
| 262K | EXL3 2.70bpw (per-expert variabl | vllm | 58.3 tok/s | — |
| 262K | EXL3 2.70bpw (per-expert variabl | vllm | 44.5 tok/s | — |
| 255709 | EXL3 2.70bpw (per-expert variabl | vllm | — | — |
this is deepseek v4.1 flash, the fastest vllm build on 2x dgx sparks, and the weights are public now 🐋 i quantized them myself from deepseek's official checkpoint, every one of the 15,360 routed experts sits at 2 to 5 bits by how sensitive it measured, 2.70 bits on average, and the whole folder is 196 gb with the two engram shards pulled straight from deepseek's own repo and a sha256 manifest that checks every file…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 21 posts from the same window are listed here in short form, each with its source post.
- B70 x2 × Qwen3.8-Flash-Next — 70.0 tok/s decode (1 measurement) · @lookoutitsbbear
- RTX 4090 x4 × GLM-5.3-Flash — 118.0 tok/s decode (3 measurements) · @PeasantSmith
- RTX 3090 x2 × MiniCPM-V 4.7-35B-A3B-int4 — 802.6 tok/s decode (6 measurements) · @Tech2Wild
- 4 DGX Sparks × GLM‑5.3 Flash — 53.0 tok/s decode (2 measurements) · @WescheNex1q
- RTX 5090 — 143.0 tok/s decode (1 measurement) · @aktsmm
- RTX 3090 × MiniCPM-V 4.7 35B-A3B — 91.0 tok/s decode (6 measurements) · @ItsmeAjayKV
- DGX Spark × GLM-5.3 — 286.0 tok/s decode (1 measurement) · @norsyx
- Spark x4 × GLM-5.3 — 32.0 tok/s decode (1 measurement) · @WescheNex1q
- DGX Spark x2 × DeepSeek-V4.1-Flash — 141.3 tok/s decode (9 measurements) · @sudoingX
- DGX Spark x2 × DeepSeek-V4.1 — 103.0 tok/s decode (1 measurement) · @sudoingX
- DGX Spark × Qwen3.8 Flash-Next — 1572.0 tok/s prefill (2 measurements) · @filicroval
- RTX 5080 16GB + RTX 5070 12GB × GLM-5.3-Flash (Maya-S24) — 177.5 tok/s decode (8 measurements) · @doc_mote
- DGX Spark x2 × DeepSeek-V4.1-Flash — 103.0 tok/s decode (7 measurements) · @sudoingX
- DGX Spark × Qwen3.8 27B Kearuga EXL3 4.00bpw — 249.6 tok/s decode (2 measurements) · @0xWhiteMage
- RTX 3090 × GLM 5.3 Flash — 53.0 tok/s decode (2 measurements) · @PeasantSmith
- Unspecified platform × Deepseek v4.1 Flash — 245.0 tok/s decode (1 measurement) · @cherry_mx_reds
- DGX Spark x2 × GLM-5.3-Flash — 210.0 tok/s decode (6 measurements) · @vr8vr8
- DGX Spark x4 × GLM-5.3 full — 34.0 tok/s decode (1 measurement) · @WescheNex1q
- RTX 6000 — 200.0 tok/s decode (1 measurement) · @volatilemarkts
- RTX 4070 Super × Qwen3.8-Flash-Next — 27.0 tok/s decode (1 measurement) · @tegamen_jp
- Unspecified platform × Deepseek-v4.1-flash — 240.0 tok/s decode (2 measurements) · @ItsmeAjayKV
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
- @0xSero — Oct 13, 2026
- @sudoingX — Oct 13, 2026
- @yuhasbeentaken — Oct 13, 2026
- @tinygrad — Oct 13, 2026
- @_nodelay — Oct 13, 2026
- @superalesha — Oct 13, 2026
- @aigclink — Oct 13, 2026
- @jmacftw — Oct 13, 2026
- @ZinkLu2 — Oct 13, 2026
- @sudoingX — Oct 13, 2026
- @sudoingX — Oct 13, 2026
- @sudoingX — Oct 13, 2026
- @lookoutitsbbear — Oct 13, 2026
- @PeasantSmith — Oct 12, 2026
- @Tech2Wild — Oct 12, 2026
- @WescheNex1q — Oct 12, 2026
- @aktsmm — Oct 12, 2026
- @ItsmeAjayKV — Oct 12, 2026
- @norsyx — Oct 12, 2026
- @WescheNex1q — Oct 12, 2026
- @sudoingX — Oct 12, 2026
- @sudoingX — Oct 12, 2026
- @filicroval — Oct 12, 2026
- @doc_mote — Oct 12, 2026
- @sudoingX — Oct 12, 2026
- @0xWhiteMage — Oct 12, 2026
- @PeasantSmith — Oct 12, 2026
- @cherry_mx_reds — Oct 12, 2026
- @vr8vr8 — Oct 12, 2026
- @WescheNex1q — Oct 12, 2026
- @volatilemarkts — Oct 12, 2026
- @tegamen_jp — Oct 12, 2026
- @ItsmeAjayKV — Oct 12, 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.