X local-model bench digest — Nov 6, 2026
As of Nov 6, 2026 — 17 posts collected from X. 43 measured runs catalogued, 17 kept to the digest only. Latest source post Nov 6, 2026 03:54 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 x4 × GLM-5.3
Nov 6, 2026 · by @jaiball7
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 256K | — | expert atlas r3.3 | 40.0 tok/s | 1000.0 tok/s |
Expert Atlas for GLM-5.3 r3.3 is out: an engine that runs GLM-5.3 on 4× DGX Spark, with no switch. • 1M context with an fp4 KV cache • On a 256K-token coding task: prefill ~1,000 tok/s, decode ~40 tok/s (speed depends on the prompt; a rough guide) • The most-used experts get more bits, raising accuracy in the same memory 1. Ring setup No 200G switch: just cable the 4 nodes in a ring with ConnectX-7. The panel's…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (quantization, hardware state, measurement method).
RTX 5060 Ti 16GB + RTX 4060 Ti 16GB × Uncensored (Strata) / Basalt official NVFP4-DQ6_K
Nov 6, 2026 · by @ren_ai_coach
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 100K | IQ4_XS | strata | — | 1447.0 tok/s |
| 100K | NVFP4-DQ6_K | basalt | — | 2436.0 tok/s |
| — | IQ4_XS | strata | 49.3 tok/s | — |
| — | IQ4_XS | strata | 48.7 tok/s | — |
| — | NVFP4-DQ6_K | basalt | 42.5 tok/s | — |
| — | Q4_K_M | basalt | 36.1 tok/s | — |
【Strata vs Basalt】の追加報告 長文処理で約40%高速化は確定 今回は4構成×各10回の反復測定を実施しました。 【検証環境】 Windows 11 / Core i7-13700F / RAM 128GB RTX 5060 Ti 16GB ＋ RTX 4060 Ti 16GB A：Strata 1GPU / Uncensored IQ4_XS B：Strata 2GPU / Uncensored IQ4_XS C：Basalt 1GPU / 公式 NVFP4-DQ6_K D：Basalt 1GPU / Uncensored Q4_K_M （AとBは同一のモデル重みを使用しています） 【① 100KトークンのTTFT】 A：69.17秒 B：74.45秒 C：41.14秒 D：63.74秒 Basalt CはStrata A比で40.5%短縮 Prefill速度もStrata Aの1,447…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state).
RTX 3090 × DeepSeek V4.1 Flash
Nov 6, 2026 · by @mapte_fireball
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | 2-bit | delta inference engine | 41.0 tok/s | — |
DeepSeek V4.1 Flash向け推論エンジン「Delta Inference Engine」が更新、RAM 64GBから動くように。 RTX 3090 1枚・RAM 192GBなら2-bit版で41 tok/s。 99問ベンチは57点で、フルモデルAPIの59点まで2点差。 64GBのPCでも手が届くのはうれしい。 数値は投稿者の測定。条件は返信に。
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, measurement method).
AMD Radeon RX 7900 XTX × Qwen3.8-27B Uncensored
Nov 6, 2026 · by @Oluwaphilemon1 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 110.0 tok/s | — |
| — | — | — | — | 1000.0 tok/s |
Qwen3.8-27B Uncensored is reportedly reaching 110 tokens/s on a single AMD Radeon RX 7900 XTX. No multi-GPU setup or dedicated inference server is required. The model runs locally on one consumer graphics card. Reported performance GPU: AMD Radeon RX 7900 XTX Model: Qwen3.8-27B Uncensored Decode speed: Up to 110 tok/s Prompt processing: Up to 1,000 tok/s Deployment: Single-GPU local inference The two speed figures…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 3060 × Qwen3.8-35B-A3B
Nov 6, 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 |
A 12GB RTX 3060 is reportedly generating around 50 tokens per second with Qwen3.8-35B-A3B. The same setup reaches approximately 550 tok/s in prompt processing, using only 16GB of system RAM alongside the GPU. No 24GB graphics card. No multi-GPU server. Just a consumer GPU, limited system memory and a carefully optimised inference stack. The experiment uses Qwen3.8-35B-A3B, an MoE model, with aggressive quantisation…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
RTX 5070 12GB x1 + RTX 5060 Ti 16GB × Qwen3.8-Flash-Next IQ3_S
Nov 5, 2026 · by @murasametech
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 131072 | IQ3_S | strata v0.2.10 | 55.9 tok/s | — |
| 262144 | IQ3_S | strata v0.2.10 | 44.7 tok/s | — |
ローカルLLMの上限を広げる前に、モデル自体が何トークンまで対応しているかを確かめました。 使ったのはQwen3.8-Flash-Next IQ3_S。モデルの情報には、もともと262,144トークンまでと書かれていて、256Kでも追加の設定は要りませんでした。エンジンの上限だけ広げても、モデルが対応していなければ意味がありません。 推論エンジンはStrata v0.2.10、構成はRTX 5070 12GB×1＋RTX 5060 Ti 16GB×1（計2枚）です。 もう一つ見たのが、VRAMの取り分です。上限を広げると、GPUに載るexpertが少し減りました。原因は確かめていません。 ・128K：5070に3,033、5060 Tiに4,321 ・256K：5070に2,946、5060 Tiに4,276 短い質問での文章を生成する速さは、128Kで55.9、256Kで44.7…
Digest-only: incomplete methodology (measurement method).
DGX Spark × GLM-5.3-Flash
Nov 5, 2026 · by @niklaslenz_ai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | project maya v1.1.53 | 16.0 tok/s | — |
| — | — | project maya v1.1.53 | 16.0 tok/s | — |
| 8192 | — | project maya v1.1.53 | — | 689.0 tok/s |
| 131072 | — | project maya v1.1.53 | 14.5 tok/s | — |
GLM-5.3-Flash, a 321B MoE, decodes 16.0 tok/s on one DGX Spark with a 128K window 1 DGX Spark, 128 GB, Project Maya v1.1.53 Decode, code: 16.0 tok/s Decode, prose: 16.0 tok/s 500 words (code and prose): about 43.2 s Prefill at 8K: 689 tok/s First token at 32K: 46.9 s First token at 128K: 193.4 s Decode with a 128K prompt: 14.5 tok/s Quant: Maya-S IQ2_XXS, 96.5 GB, about 2.4 bpw All 12,096 experts fit in the Spark's…
Full methodology stated.
MI350X x8 × Kimi K3
Nov 5, 2026 · by @net_termina · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | tinygrad | — | 38.6 tok/s |
| — | — | tinygrad | 6.2 tok/s | — |
| — | — | tinygrad | — | 43.0 tok/s |
| — | — | tinygrad | 9.7 tok/s | — |
Someone shared a post on 4xR9700 Tinybox that claims to run Kimi K3 for $30k at 40tok/s with good prefill. Whenever I see something that sounds way too good to be true, I cannot resist doing my due diligence. First, I wish them all the success and think what they are doing is amazing to make large models more accessible. However, I can’t fathom announcing a product without every single possible spec/bench to back up…
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
Nov 5, 2026 · by @Picomoorhead
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | pico-dgx-ds4 (fork of blaesus/ds | 80.0 tok/s | — |
| — | — | pico-dgx-ds4 (fork of blaesus/ds | 138.0 tok/s | — |
| — | — | pico-dgx-ds4 (fork of blaesus/ds | 438.0 tok/s | — |
| — | — | pico-dgx-ds4 (fork of blaesus/ds | — | 10000.0 tok/s |
| — | — | sglang | 64.0 tok/s | — |
| — | — | sglang | 117.0 tok/s | — |
| — | — | sglang | 427.0 tok/s | — |
| — | — | sglang | — | 5300.0 tok/s |
438 tok/s at 16 users DeepSeek V4.1 Flash on 4x DGX Spark, now open source: 80 chat / 138 code single stream 438 tok/s at 16 users 1M context, 10k tok/s prefill image support soon https://github.com/Picom-code/pico-dgx-ds4 https://t.co/nC2perdyOZ
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state).
DGX Spark x2 × GLM Flash
Nov 5, 2026 · by @Wappy_x
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | FP4 | vade | 2.3 tok/s | — |
| — | FP4 | vade | 37.8 tok/s | — |
Earlier today GLM just had its first heartbeats on Vade Vade is an inference engine I've been building for the NVIDIA DGX Spark. It's all Rust, with no third-party crates, and it runs on its own GPU kernels. The whole point is that the same request gives you the same answer every time, whether it runs on 1, 2 or 4 Sparks, and whether your chats are processed alone, batched or cached. Two days ago GLM Flash wouldn't…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, measurement method).
RTX 5090 × Clover-35B-A3B
Nov 5, 2026 · by @witcheer
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | Q4_K_M | llama.cpp PR build | 88.6 tok/s | — |
most language models write one token at a time. Clover-35B-A3B, a block-diffusion model from DiffuSpace, guesses several tokens in one pass, then checks them and keeps the ones the check agrees with. I gave it the same Mont-Saint-Michel prompt as the rest of the series, one RTX 5090 and Hermes Agent's /goal. - it passed our gate in 13 minutes (the series gives each model an hour) - the /goal judge marked it done…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
Unspecified platform × Strata
Nov 5, 2026 · by @tegamen_jp
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata | 50.0 tok/s | — |
Strataの最新版 VRAM 6GB で 50 tok/s エグいwww ー VRAMとRAMは別々に管理 ー VRAM 上級設定から簡単に上限を指定 RAM 自動的に余力を残した値になる JSONの書き換えで好きな上限にもできそう 次は、ComfyUI、Unity、Blenderなどを 同時起動で検証してみる https://t.co/LixRXaBLoj
Digest-only: confidence 0.55 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 5 posts from the same window are listed here in short form, each with its source post.
- DGX Spark x2 × GLM-5.3-Flash — 58.8 tok/s decode (4 measurements) · @sfxnz
- RTX 5090 — 480.0 tok/s decode (1 measurement) · @coffeecup2020
- Colab G4 × DeepSeek-V4.1 — 52.0 tok/s decode (1 measurement) · @Delroy715
- Mac Studio M5 Max × Qwen3.5-27B — 26.0 tok/s decode (2 measurements) · @AlexObuchowski
- RTX 3060 × Qwen3.8-35B-A3B — 50.0 tok/s decode (1 measurement) · @LiLensss
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
- @jaiball7 — Nov 6, 2026
- @ren_ai_coach — Nov 6, 2026
- @mapte_fireball — Nov 6, 2026
- @Oluwaphilemon1 — Nov 6, 2026
- @Oluwaphilemon1 — Nov 6, 2026
- @murasametech — Nov 5, 2026
- @niklaslenz_ai — Nov 5, 2026
- @net_termina — Nov 5, 2026
- @Picomoorhead — Nov 5, 2026
- @Wappy_x — Nov 5, 2026
- @witcheer — Nov 5, 2026
- @tegamen_jp — Nov 5, 2026
- @sfxnz — Nov 5, 2026
- @coffeecup2020 — Nov 5, 2026
- @Delroy715 — Nov 5, 2026
- @AlexObuchowski — Nov 5, 2026
- @LiLensss — Nov 5, 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.