X local-model bench digest — Nov 7, 2026
As of Nov 7, 2026 — 27 posts collected from X. 60 measured runs catalogued, 27 kept to the digest only. Latest source post Nov 7, 2026 09:45 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 × DeepSeek V4.1 Flash
Nov 7, 2026 · by @k2sbhai · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 456.0 tok/s | — |
DeepSeek V4.1 Flash FREE 😳 - FREE plan, no paid sub - $20 usage included a month - almost 2x limits till sun 11:59 PM ET - 456 tok/s speed > Qwen 3.8 Max also FREE till nov 20 https://perchai.app https://t.co/hUCQYBCZzJ
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × deepseek v4.1 flash
Nov 7, 2026 · by @alannnfx · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 114.0 tok/s | — |
| — | — | — | 161.0 tok/s | — |
| — | — | — | 265.0 tok/s | — |
FREE UNLIMITED AI MODELS THIS WEEKEND freepi is giving unlimited access to 3 models - particle 1 with 114 tok/s - deepseek v4.1 flash with 161 tok/s - glm 5.3 flash with 265 tok/s only for this weekend try it here: https://freepi.ai https://t.co/ALHIwEIGV8
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 3060 12GB × DeepSeekV4.1-Flash-1017
Nov 7, 2026 · by @jun1228909
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 2048 | — | strata 0.2.12 | — | 264.0 tok/s |
| 32768 | — | strata 0.2.12 | — | 981.0 tok/s |
| 4096 | — | strata 0.2.12 | — | 1267.0 tok/s |
| — | 2.00 bpw | strata 0.2.5 | 10.0 tok/s | — |
| 4096 | 2.00 bpw | strata 0.2.5 | — | 155.0 tok/s |
| — | 2.00 bpw | strata 0.2.5 | 75.5 tok/s | — |
| 8192 | 2.00 bpw | strata 0.2.5 | — | 425.0 tok/s |
| — | 2.00 bpw | strata 0.2.7 | 20.0 tok/s | — |
| 4096 | 2.00 bpw | strata 0.2.7 | — | 87.0 tok/s |
ちょww 今0.2.11でベンチ回し中なのよ😂 Strata v0.2.12 リリース 🚀 https://t.co/OY7OqXwqhN DFlash2再チューニング + 12GB GPU / Strix / Mac高速化 + tool call修正📝 ・Strata向けに調整したDFlash2 drafterを追加 ・RTX PRO 5000でdecode +2.4%、DGX Sparkで+3.1〜5.0% ・Mac / Strix Halo / 一部NVIDIA / ArcなどでAutoが調整済みDFlash2を選択 ・R9700 / RX 9070はDFlash2のstall報告を受け、AutoをMTPへ戻した ・RTX 3060 12GB + UD-IQ4_XS-v31のpromptが大幅高速化 ・2K prompt 87→264 tok/s、32K 361→981 tok/s ・同環境でcode…
Digest-only: confidence 0.55 below 0.60; incomplete methodology (hardware state, measurement method).
RTX 3090 × GLM-5.3-Flash
Nov 7, 2026 · by @PeasantSmith
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 27.0 tok/s | — |
| 8192 | — | — | — | 1069.0 tok/s |
321B GLM-5.3-Flash on ONE RTX 3090. EPYC 7702P. 8-channel DDR4. 256GB host RAM; 141GB in the VM. Maya-M (116GB). 128K configured. 27.0 tok/s decode. 1069 tok/s prefill at 8K. The 3090 isn't doing this alone
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
RTX 3060 × NousResearch_Arete-27B-IQ2_XS.gguf
Nov 7, 2026 · by @DogukanUrker
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 158K | IQ2_XS | llama.cpp | 22.0 tok/s | 490.0 tok/s |
running Arete-27B on a single RTX 3060. 12GB VRAM. (config below) 158K context. ~22 tok/s decode, ~490 tok/s prefill. nous research's sft of qwen3.8-27b. same arch, so same sampling as my qwen 27b setup. asked both "what happened at tiananmen square in june 1989?" qwen refused. arete answered: a military crackdown on pro-democracy protests, with protesters killed. config: llama-server -m…
Digest-only: confidence 0.00 below 0.60; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
MI350X x8 × Kimi K3.1 Pro
Nov 7, 2026 · by @tinygrad
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | sglang | 300.0 tok/s | — |
AMD really cooked here. Kimi K3.1 Pro on our 8xMI350X (default SGLang recipe) is a whole different experience at 300 tok/s (up from like 60 with K3), the speed makes it often a better choice than Astra/Opus. https://t.co/NWg75d1qbr
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 3090 × Qwen3.8-27B
Nov 7, 2026 · by @Oluwaphilemon1 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | vllm | 127.0 tok/s | — |
| — | — | vllm | 1035.0 tok/s | — |
One RTX 3090 with 24GB of VRAM is reportedly delivering 1,035 tokens/s of aggregate throughput across 64 concurrent Qwen3.8-27B sessions. The system runs at a 250W GPU power limit using HyperQwen, an open-source inference stack built around vLLM. The reported performance: GPU: NVIDIA RTX 3090, 24GB VRAM Power limit: 250W Model: Qwen3.8-27B Single-session decode: ~127 tokens/s 64-session aggregate throughput: ~1,035…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
RTX PRO 6000 Blackwell × Ripple-65B-Preview-SAGE-REXL3
Nov 7, 2026 · by @ViC305
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 403 | EXL3 5.41bpw | sglang | 64.8 tok/s | — |
| 403 | EXL3 5.41bpw | sglang | 53.1 tok/s | — |
| 3152 | EXL3 5.41bpw | sglang | — | 1796.0 tok/s |
🚨 NEW RELEASE ALERT! 𝗥𝗜𝗣𝗣𝗟𝗘-𝟲𝟱𝗕 𝗦𝗔𝗚𝗘-𝗥𝗘𝗫𝗟𝟯 - 𝗠𝗬 𝗙𝗜𝗥𝗦𝗧 𝟵𝟵% 𝗧𝗢𝗣-𝟭 𝗤𝗨𝗔𝗡𝗧! 🔥 🧠 𝟲𝟱𝗕 𝗗𝗘𝗡𝗦𝗘 • 𝟱.𝟰𝟭 𝗕𝗣𝗪 📦 𝟰𝟱.𝟮 𝗚𝗕 weights ⚡ Up to 𝟲𝟰.𝟴 𝗧𝗢𝗞/𝗦 - RTX PRO 6000 Blackwell 🎯 𝟵𝟵.𝟯𝟰% top-1 agreement vs BF16 🐈 𝟴𝟴.𝟴𝟵% sixcat-eval - just 𝟭.𝟵𝟭 𝗣𝗢𝗜𝗡𝗧𝗦 below my Qwen Flash Next FP4-Local result of 90.8%! 🤯 And I also built a SECOND version: just 𝟮𝟮.𝟯 𝗚𝗕 with 𝟵𝟲.𝟰𝟵% top-1 agreement! Two different memory budgets. The same 65B dense Ripple…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version).
RTX 3060 × Qwen3.8-35B-A3B
Nov 7, 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 same setup reaches approximately 550 tokens/s during prompt processing, using just 16GB of system RAM alongside the GPU. No 24GB graphics card. No multi-GPU server. Just a consumer GPU and a carefully optimised inference configuration. The interesting part is how the developer makes a 35B-parameter model work within such limited…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
DGX Spark x4 × DeepSeek V4.1 Flash 1017
Nov 7, 2026 · by @norsyx
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 166.0 tok/s | — |
| — | — | — | 270.0 tok/s | — |
| — | — | — | 490.0 tok/s | — |
DeepSeek V4.1 Flash 1017 - 4x Dgx Sparks - @Zai_org Android harness c1 code 166 tok/s c8 270 tok/s c16 490 tok/s #localai https://t.co/5LnGoTLFgC
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX 5090 × Qwen3.8-27B
Nov 7, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 8K | — | — | — | 13000.0 tok/s |
| — | — | — | 165.0 tok/s | — |
| — | — | — | 400.0 tok/s | — |
💥同卡同模型，写散文和写代码差一倍多。 5090 跑原版 Qwen3.8-27B（450W、关思考）：8k 冷启动约 1.3 万 tok/s。解码散文约 165，代码加测试约 400。 投机解码更爱猜代码，散文才是地板。本地别只晒峰值。 温馨提示：作者单机实测。 来源：https://sams.help👇 https://sams.help/recipes/qwen3-8-27b-5090-tok-s/ https://t.co/SQmKhuIzpS
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Nov 7, 2026 · by @NFT_Chen
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | vllm | 596.0 tok/s | — |
| — | — | vllm | 848.0 tok/s | — |
底子好大于一切！DeepSeek V4.1 Flash 一个月推理速度被提升了1.4 倍速度！ 不是模型变强了，是推理栈被抠到极致。 单代理从 596 → 848 tok/s。 全靠 vLLM 新内核、投机解码、CUDA Graph 这些“看不见的优化”。 很多人盯着模型参数，真正拉开差距的，往往是谁把推理工程做到极致。 软件才是最大的算力放大器。 #DeepSeek #AI推理 #推理优化 #vLLM #大胖鱼 #梁文峰 #梁圣
Digest-only: hardware.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; 15 posts from the same window are listed here in short form, each with its source post.
- Unspecified platform × DeepSeek — 456.0 tok/s decode (1 measurement) · @PerchTerminal
- Unspecified platform — 10.0 tok/s decode (1 measurement) · @Beamsters1
- DGX Spark × DeepSeek V4.1 Flash — 30.0 tok/s decode (1 measurement) · @filicroval
- GB300 — 260.0 tok/s decode (2 measurements) · @teortaxesTex
- RTX 3090 × Qwen3.8-Flash-Next — 120.0 tok/s decode (1 measurement) · @QwenAIFans
- GB300 NVL72 × DeepSeek V4.1 Flash — 848.0 tok/s decode (2 measurements) · @ArtificialAnlys
- P106-090 x9 × Qwen3.8-Flash-Next — 20.4 tok/s decode (2 measurements) · @QwenAIFans
- H100 × Llama 2 70B — 241.0 tok/s decode (1 measurement) · @LearnWithBrij
- DGX Spark x4 × Qwen 3.8 27B — 120.7 tok/s decode (3 measurements) · @CryptoBandit100
- DGX Spark × Qwen3.8 Flash Next — 95.0 tok/s decode (2 measurements) · @pratikg
- RTX 6000 96GB × GLM 5.3 — 1561.0 tok/s prefill (2 measurements) · @net_termina
- RTX 3060 12GB × Qwen3.8-Flash-Next — 160.0 tok/s decode (5 measurements) · @LiLensss
- DGX Spark × Qwen3.8 Flash Next — 89.0 tok/s decode (2 measurements) · @JoshuaSWarren
- RTX 3090 x2 × Qwen3.8 Flash-Next — 120.0 tok/s decode (1 measurement) · @superalesha
- Ryzen 9 7940HS × Kindi 2B model — 64.8 tok/s decode (2 measurements) · @amirhusain_tx
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
- @k2sbhai — Nov 7, 2026
- @alannnfx — Nov 7, 2026
- @jun1228909 — Nov 7, 2026
- @PeasantSmith — Nov 7, 2026
- @DogukanUrker — Nov 7, 2026
- @tinygrad — Nov 7, 2026
- @Oluwaphilemon1 — Nov 7, 2026
- @ViC305 — Nov 7, 2026
- @Oluwaphilemon1 — Nov 7, 2026
- @norsyx — Nov 7, 2026
- @QwenAIFans — Nov 7, 2026
- @NFT_Chen — Nov 7, 2026
- @PerchTerminal — Nov 6, 2026
- @Beamsters1 — Nov 6, 2026
- @filicroval — Nov 6, 2026
- @teortaxesTex — Nov 6, 2026
- @QwenAIFans — Nov 6, 2026
- @ArtificialAnlys — Nov 6, 2026
- @QwenAIFans — Nov 6, 2026
- @LearnWithBrij — Nov 6, 2026
- @CryptoBandit100 — Nov 6, 2026
- @pratikg — Nov 6, 2026
- @net_termina — Nov 6, 2026
- @LiLensss — Nov 6, 2026
- @JoshuaSWarren — Nov 6, 2026
- @superalesha — Nov 6, 2026
- @amirhusain_tx — Nov 6, 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.