X local-model bench digest — Nov 5, 2026
As of Nov 5, 2026 — 37 posts collected from X. 99 measured runs catalogued, 37 kept to the digest only. Latest source post Nov 5, 2026 10:07 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.
RTX 3060 × Qwen3.8-Flash-Next
Nov 5, 2026 · by @ai_hakase_ · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | rustingstrata | 36.0 tok/s | 370.0 tok/s |
| — | — | rustingstrata | 113.0 tok/s | — |
【Pure Rust製推論エンジンRustingStrataで巨大MoEを高速化！】 VRAMに収まりきらないQwenの巨大MoEモデルを低VRAM環境で効率よく動かす、Pure Rust製の推論エンジン「RustingStrata」が登場しました！ RTX 3060（VRAM 12GB）の検証環境において、Qwen3.8-Flash-Nextのデコード速度が約36 tok/s、プリフィル速度が平均約370 tok/s（最大約610 tok/s）を記録していますね。 最大の特徴は、VRAMに入り切らないエキスパートをCPU側で動作させつつ、頻繁に使われるものをGPUへ動的に移動（プロモーション）させるハイブリッド推論メカニズムです！ さらに、C++のビルドを排除し、起動時のNVRTCによるCUDAカーネルコンパイルやVulkan（SPIR-V）を採用しているのも面白いポイントですね。…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × Qwen3.8-Flash-Next
Nov 5, 2026 · by @NegiTurkey
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 128K | IQ3_S | strata | 236.9 tok/s | — |
後回しにしていたStrataを、ちもろぐさんの分かりやすい解説を読みながら導入してみました。 躓きそうな部分が細かく書いてあるので導入・設定がやりやすいですね！ 最近はProject-Mayaや170HX専用レシピを使っており、Strataで同様のモデルを入れても容量を食うだけかもと思っていましたが、Windows版をインストールしてみると非常にユーザーフレンドリーで何をすれば良いのかが分かりやすい！ ・・・なるほどこれは評判が良いのも納得！ 当初衝撃的なデビューを飾ったAnti◯ravityなんて、未だにパッチ入れないと英語しか選べない不親切＆閉鎖的仕様だし...。 Strataで最初に試したのはQewen3.8-Flash-Next IQ3_S 128Kで、モデルの自動調整を試したところ236.9 tok/sで調整不要と表示。 Deepseek…
Digest-only: confidence 0.50 below 0.60; hardware.name missing; incomplete methodology (engine+version, hardware state, measurement method).
RTX 5090 × Qwen3.8-Flash-Next
Nov 5, 2026 · by @jun1228909
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | NVFP4 | strata | 197.0 tok/s | — |
| — | — | strata | 145.0 tok/s | — |
| — | — | strata | 176.0 tok/s | — |
| — | — | strata | 187.0 tok/s | — |
| — | — | strata | 251.0 tok/s | — |
| — | — | strata | 279.0 tok/s | — |
| — | — | strata | 343.0 tok/s | — |
自作ベンチ strata-jbench のご紹介🪛 https://t.co/LfHzfxfyHx とりあえず RTX5090/DDR5 128GB/Ubuntu、ハーネスなしのベースラインで Qwen FN 系 18 モデル分回してみた結果 https://t.co/QUvV5Bvm5K よく試してる「マイクラ風3D箱庭ゲーム作って」のプロンプト、いい加減再現性のある形で管理したいなと思って始めました😂 jbench には Tier3 のタスクとして混ぜてあります ただ、現状は「いろんな観点で tok/s を取ってみました集」という感じなので、精度（正答率）の評価をどうしようかなぁというところ 以下、最初の気づきメモ📝 （誰か詳しい方…🥹） ■ Tier1（各モデルの素の速さ観点） 👉 大きさのわりに速い NVFP4 ・128 GB の NVFP4（197 tok/s）は、111 GB の unsloth…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state).
DGX Spark x2 × GLM-5.3-Flash
Nov 5, 2026 · by @atasrb
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | FP4 | tensorfold 1.1.9 | 46.0 tok/s | — |
GLM-5.3-Flash on two DGX Sparks built the same live monitor in 6 minutes, from the same prompt I gave Qwen3.8-Flash-Next yesterday. The setup: NVIDIA's FP4 build on TensorFold 1.1.9, TP=2 over the QSFP link, opencode 2.2.5 with all permissions on, inside a Docker sandbox, thinking off. 13.2K tokens at about 46 tok/s while writing. The video runs in real time for 12 s in the middle, where a typo breaks App.jsx and…
Digest-only: confidence 0.00 below 0.60; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark × Ripple-65B Preview
Nov 5, 2026 · by @aijoey
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 128K | NVFP4 | sglang | 20.5 tok/s | — |
| 128K | NVFP4 | sglang | 26.6 tok/s | — |
| 128K | NVFP4 | sglang | 7.6 tok/s | — |
Got Ripple-65B Preview by @RicursiveAI running on one DGX Spark w/Hermes agent. Congrats on the release, @Azaliamirh and @annadgoldie. NVFP4 weights, Ricursive’s BF16 draft for DFlash2 (@zhijianliu), and 128K context configured. Hooked it into DeepSeek Harness with streaming and tool calls working. Across three measured runs per prompt: Code: 20.5 tok/s Edits: 26.6 tok/s Prose: 7.6 tok/s Those are end to end rates…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state).
G4 96GB × DeepSeek-V4.1
Nov 5, 2026 · by @Delroy715
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 52.0 tok/s | — |
别再让 Google AI Ultra 的 1000 CCU 吃灰了！ 实测用单张 96GB G4 显卡把 748B DeepSeek-V4.1 变成不限速私人 API：165GB 骨干全进显存+内存，52 tok/s 满速流式，一套权重同时刷出标准/无审查双模型，还顺着扒出了 196B Engram 记忆表的底层真相，全套一键脚本已开源👇 https://x.com/i/article/2117841043337138176
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
RTX 3090 × Qwen3.8-27B
Nov 5, 2026 · by @Oluwaphilemon1 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | vllm | 127.0 tok/s | — |
| — | — | vllm | 1035.0 tok/s | — |
One RTX 3090 is reportedly delivering over 1,000 tokens/s while serving 64 concurrent Qwen3.8-27B sessions. The GPU has just 24GB of VRAM and runs with a 250W power limit. No multi-GPU setup or dedicated inference accelerator is involved. The project is called HyperQwen, an open-source inference setup built around vLLM. Reported performance • GPU: NVIDIA RTX 3090, 24GB VRAM • Power limit: 250W • Model: Qwen3.8-27B •…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, measurement method).
RTX 3060 × Qwen3.8-35B-A3B
Nov 5, 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/s with Qwen3.8-35B-A3B. The setup also reaches approximately 550 tokens/s in prompt processing, using just 16GB of system RAM alongside the GPU. No 24GB graphics card or multi-GPU server is involved. The experiment relies on MoE architecture, aggressive quantization and hybrid inference. The hardware and configuration GPU: NVIDIA RTX 3060, 12GB VRAM System…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
Unspecified platform × Qwen
Nov 5, 2026 · by @Multi_Serio_Ai · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 200.0 tok/s | — |
Strata의 구조로 로컬에서 200 Tok/s 로 돌아가는 Qwen 4.0 Flash next와 Glm 5.5 Flash 가 있으면 정말 재미있을 것 같다. 그리고 아마 올 연말 ~ 내년 초엔 그리 되지 않을까. 지능은 생각보다 더 저렴해지고 더 빨라진다. 연초까지도 이런건 생각조차 못했다. 하지만 사람이 거기에 따라갈 수 있는지는 생각해 볼 문제다.
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 5090 × Qwen3.8
Nov 5, 2026 · by @Xipzer
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | infernix | 200.0 tok/s | — |
Local AI is actually exploding. Never been so grateful for buying something I genuinely didn’t need (RTX 5090 at MSRP) 🤣 Infernix is a game changer - will contribute if/when I have time. 200+ tok/s powering Wendy basically feels like realtime. Highly suggest people move over to this for now if they’re using Qwen3.8 on local inference. https://t.co/dNRMqcbyGd
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 1017
Nov 5, 2026 · by @isoquant_labs
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 600.0 tok/s | — |
DeepSeek V4.1 Flash 1017 runs at 600+ tok/s on Isoquant Ships Tetris in 10 seconds, 8.7x faster than Claude Opus 5.5. This is our AI optimizing AI, and we are passing performance wins and cost savings on to our customers: → 0.24s P50 / 1.16s P95 time to first token → $0.035 input / $0.005 cached / $0.37 output per 1M tokens Keep your code and harness, just change the base URL. Works with your existing OpenAI and…
Digest-only: hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform
Nov 5, 2026 · by @Schnitzel · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 1500.0 tok/s | — |
what people plugged into 1500 tok/s DeepSeek this week: opencode, deepseek-harness, ZCode, Claude Code, Codex, pi, and a lot of their own scripts what are you building with it?
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).
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.
- RX 5700 XT × TwIL-8B — 43.5 tok/s decode (1 measurement) · @GuideboardLabs
- Unspecified platform × DeepSeek 4.1 Flash — 20.0 tok/s decode (1 measurement) · @MaximeRivest
- RTX 6000 × GLM 5.3 753B — 45.0 tok/s decode (2 measurements) · @net_termina
- H200 x4 × deepseek-v4.1-flash-1017 — 8100.0 tok/s decode (2 measurements) · @0xSero
- DGX Spark × Gemma 4 26B A4B — 60.6 tok/s decode (2 measurements) · @Davidstout
- B300 x8 × Kimi K3.1 Pro — 100.0 tok/s decode (3 measurements) · @MrCatid
- V100 32G x2 × Qwen3.8-27B-QSA — 45.0 tok/s decode (5 measurements) · @LiLensss
- EPYC server × strata — 190.0 tok/s decode (3 measurements) · @shutoito
- RTX 5090 × Nanbeige4.3-33B — 78.9 tok/s decode (2 measurements) · @witcheer
- DGX Spark × GLM-5.3 Flash — 40.2 tok/s decode (10 measurements) · @filicroval
- Unspecified platform × Claude Opus 5.5 — 111.0 tok/s decode (2 measurements) · @TeksEdge
- Unspecified platform × Qwen3.8 27B QSA GSQ-RCO IQ3_S — 115.0 tok/s decode (2 measurements) · @pentacoxian_dev
- DGX Spark x2 × GLM-5.3-Flash — 87.7 tok/s decode (6 measurements) · @niklaslenz_ai
- M5 Ultra — 15.0 tok/s decode (1 measurement) · @fawnunwrap
- RTX 6000 PRO × Qwen3.8-Flash-Next — 170.0 tok/s decode (1 measurement) · @NFT_Chen
- M5 Max × Qwen 3.8 27B — 1727.0 tok/s decode (1 measurement) · @Leik0w0
- VR200 × strongest open model yet — 800.0 tok/s decode (1 measurement) · @mgoin_
- Unspecified platform × GPT-6 Astra — 111.0 tok/s decode (3 measurements) · @ArtificialAnlys
- B200 — 1881.0 tok/s decode (2 measurements) · @mimi0xyz
- DGX Spark × Deepseek-4.1-Flash-1017 — 40.0 tok/s decode (1 measurement) · @ViC305
- GB300 × Kimi K3.1 Pro — 33700.0 tok/s decode (2 measurements) · @sgl_project
- RTX 5090 × Qwen3.8-27B — 1260.0 tok/s decode (10 measurements) · @LiLensss
- DGX Spark x2 × GLM 5.3 Flash — 87.0 tok/s decode (3 measurements) · @MiaAI_lab
- DGX Spark x2 × DeepSeek-V4.1-Flash-1017 — 249.0 tok/s decode (7 measurements) · @bertholomusai
- MSI Prestige N16 Flip AI Plus (RTX Spark N1X) × GPT-OSS-120B — 65.4 tok/s decode (3 measurements) · @fishjar
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
- @ai_hakase_ — Nov 5, 2026
- @NegiTurkey — Nov 5, 2026
- @jun1228909 — Nov 5, 2026
- @atasrb — Nov 5, 2026
- @aijoey — Nov 5, 2026
- @Delroy715 — Nov 5, 2026
- @Oluwaphilemon1 — Nov 5, 2026
- @Oluwaphilemon1 — Nov 5, 2026
- @Multi_Serio_Ai — Nov 5, 2026
- @Xipzer — Nov 5, 2026
- @isoquant_labs — Nov 5, 2026
- @Schnitzel — Nov 5, 2026
- @GuideboardLabs — Nov 5, 2026
- @MaximeRivest — Nov 5, 2026
- @net_termina — Nov 4, 2026
- @0xSero — Nov 4, 2026
- @Davidstout — Nov 4, 2026
- @MrCatid — Nov 4, 2026
- @LiLensss — Nov 4, 2026
- @shutoito — Nov 4, 2026
- @witcheer — Nov 4, 2026
- @filicroval — Nov 4, 2026
- @TeksEdge — Nov 4, 2026
- @pentacoxian_dev — Nov 4, 2026
- @niklaslenz_ai — Nov 4, 2026
- @fawnunwrap — Nov 4, 2026
- @NFT_Chen — Nov 4, 2026
- @Leik0w0 — Nov 4, 2026
- @mgoin_ — Nov 4, 2026
- @ArtificialAnlys — Nov 4, 2026
- @mimi0xyz — Nov 4, 2026
- @ViC305 — Nov 4, 2026
- @sgl_project — Nov 4, 2026
- @LiLensss — Nov 4, 2026
- @MiaAI_lab — Nov 4, 2026
- @bertholomusai — Nov 4, 2026
- @fishjar — Nov 4, 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.