X local-model bench digest — Nov 1, 2026
As of Nov 1, 2026 — 22 posts collected from X. 64 measured runs catalogued, 22 kept to the digest only. Latest source post Nov 1, 2026 11:17 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 3090 x4 × DeepSeek V4.1 Flash
Nov 1, 2026 · by @superalesha
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 67.7 tok/s | — |
| — | — | — | 57.5 tok/s | — |
| — | — | — | 46.5 tok/s | — |
I do not recommend HUANANZHI motherboards to anyone. Especially with an EPYC. I bought 6 more 64 GB sticks for my H12D-8D, so DeepSeek V4.1 Flash could run faster on my 4x 3090. 2 days later I am back to 4 sticks and 256 GB. 8 sticks: hangs on code Ad (wtf?). With BIOS tweaks it finally boots, sees 7 of them, and only at 1866. 6 sticks: boots, but without a fan pointed at them they overheat in a few minutes and…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, measurement method).
RTX 5090 × Qwen3.8-27B
Nov 1, 2026 · by @niklaslenz_ai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | tensorfold 1.2.0 | 268.6 tok/s | — |
| — | — | tensorfold 1.2.0 | 248.6 tok/s | — |
| 32768 | — | tensorfold 1.2.0 | — | 3954.0 tok/s |
| — | — | tensorfold 1.2.0 | 211.8 tok/s | — |
| — | — | tensorfold 1.1.2 | 268.2 tok/s | — |
| — | — | tensorfold 1.1.2 | 238.7 tok/s | — |
| 32768 | — | tensorfold 1.1.2 | — | 4283.0 tok/s |
| — | — | tensorfold 1.1.2 | 204.6 tok/s | — |
TensorFold 1.2.0 decodes Qwen3.8-27B code at 268.6 tok/s on an RTX 5090, level with 1.1.2. Prefill is 5.5 to 7.7% lower. RTX 5090 (32 GB), Qwen3.8-27B on TensorFold 1.2.0: Decode, code: 268.6 tok/s Decode, prose: 248.6 tok/s Prefill, 32K: 3,954 tok/s First token, 32K: 8.29 s 8 at once: 211.8 tok/s RTX 5090 (32 GB), Qwen3.8-27B on TensorFold 1.1.2 (Oct 24): Decode, code: 268.2 tok/s Decode, prose: 238.7 tok/s…
Full methodology stated.
RTX 3090 + 80GB × Qwen3.8 Flash Next
Nov 1, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata | 100.0 tok/s | 1473.0 tok/s |
💥 3090 + 80GB，Qwen3.8 Flash Next 写到 100 tok/s。 Strata 本地：读提示约 1473 tok/s，生成约 100。作者还在问有没有更快的。 温馨提示：作者自测，上下文长度未写清。你觉得还能再快吗？ 来源：@cheatyyyy👇 https://t.co/fo2BzPTje9
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RTX PRO 6000 x2 × GLM 5.3 Flash
Nov 1, 2026 · by @MiaAI_lab
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | EXL3 4 bpw | tensorfold | 245.0 tok/s | — |
| — | EXL3 4 bpw | tensorfold | 336.0 tok/s | — |
| — | EXL3 4 bpw | tensorfold | 436.0 tok/s | — |
| 64K | EXL3 4 bpw | tensorfold | — | 6000.0 tok/s |
| 128K | EXL3 4 bpw | tensorfold | — | 5785.0 tok/s |
| 256K | EXL3 4 bpw | tensorfold | — | 5503.0 tok/s |
Run GLM 5.3 Flash on your @NVIDIAAI 2x RTX PRO 6000s with ease, on TensorFold 🔥 These cards are monsters, and it shows. It uses the same 4 bpw EXL3 quant as on the DGX Sparks. All numbers below are at 400W. ABLIT flag included (gated)!!! Performance (non-greedy, thinking on): - 245 tok/s on a single stream - 336 tok/s on 2 concurrent streams - 436 tok/s on 4 concurrent streams Prompt processing (prefill): - 4k-64k…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
RTX 5070 × DeepSeek V4.1 (uncensored)
Nov 1, 2026 · by @alizeh_ai
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 5.0 tok/s | — |
what the f*ck, an uncensored 18+ DeepSeek V4.1 just dropped😳 and it runs on a 12GB graphics card - original says no to 85.2% of harmful test prompts, this one says no to 0% - even with thinking on, refusals go from 27 out of 30 to 0 - needs 12GB GPU, about 64GB RAM and 257GB free SSD - slow, about 4 to 6 tokens per second on an RTX 5070 - free, MIT license try it here: https://t.co/DEy7eAm0Sy
Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
M3 Ultra × Qwen3.8-Flash
Nov 1, 2026 · by @ivanfioravanti
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 16K | — | tensorfold | 191.0 tok/s | — |
| 16K | — | tensorfold | 140.0 tok/s | — |
| 16K | — | tensorfold | 81.0 tok/s | — |
As promised I tested 8 users at once on an M3 Ultra vs single and dual DGX Spark, with Qwen3.8-Flash, 16K prompts and the TensorFold engine and @MiaAI_lab recipe for 2 x DGX. Once all 8 are generating, the Mac is the fastest: 191 tok/s, vs 140 for 2 Sparks and 81 for 1 Spark. But there's a big catch: the Mac reads the prompts one at a time. The 8th user waits ~9 minutes for the first word. On a single Spark it's…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state).
B300 × DeepSeek-V4.1-Flash
Nov 1, 2026 · by @EzoLabsInc
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 1064.0 tok/s | — |
| — | — | — | 870.0 tok/s | — |
>1,000 tok/s on DeepSeek-V4.1-Flash for one user, about double in five days. Fastest bench request: 1,064. Bench average: 870. https://ezolabs.io/research/megakernel-b300?utm_source=x&utm_medium=organic_social&utm_campaign=research-megakernel-b300&utm_content=post
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
RX 5700 XT × Strata
Nov 1, 2026 · by @GuideboardLabs
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 4096 | — | strata | 10.1 tok/s | — |
| 16384 | — | strata | 9.9 tok/s | — |
| 32768 | — | strata | 9.4 tok/s | — |
| — | — | strata | 1.9 tok/s | — |
Strata is serving a 58.4 GB model on an 8 GB card from 2019 Not a fraction offloaded. 22.4 GiB of experts page-locked in RAM, up to 685 of them resident in the 5700 XT, the rest mmap'd so the driver never tries to pin a 23 GiB arena into a 15.6 GiB GTT window. That is the failure everyone on RDNA1 knows: gfxhub page fault, GPU reset, desktop freezes. Zero page faults. Zero resets. At 4K, 16K and 32K context. On Agon…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, measurement method).
RTX 5070 12GB + RTX 5060 Ti 16GB × Qwen3.8-Flash-Next
Nov 1, 2026 · by @murasametech
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 14K | IQ3_S | strata v0.2.6 | — | 456.5 tok/s |
| 14K | IQ3_S | strata v0.2.6 | — | 108.1 tok/s |
| 14K | IQ3_S | strata v0.2.6 | — | 74.1 tok/s |
| — | IQ3_S | strata v0.2.6 | 60.0 tok/s | — |
GPU 2枚で動かしているローカルLLMを、1枚だけで動かすとどうなるか。差は想像以上でした。 Strata v0.2.6で、Qwen3.8-Flash-Next IQ3_Sを3つの構成で動かしました。RAMは64GBです。 Prefill（約1.4万tokens、各3回の中央値） ・RTX 5070 12GB×1＋RTX 5060 Ti 16GB×1（計2枚）：456.5 tok/s ・RTX 5060 Ti 16GBだけ：108.1 tok/s ・RTX 5070 12GBだけ：74.1 tok/s Decodeも、2枚の約60 tok/sに対して、1枚では28〜38 tok/sでした。…
Full methodology stated.
DGX Spark × qwen3.8-27b
Oct 31, 2026 · by @0xBakeer
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 49.9 tok/s | — |
| — | — | — | 13.7 tok/s | — |
a year ago a 27b model on one desk box meant waiting. now qwen3.8-27b does 49.89 tok/s on a single dgx spark in my bench, and plain decoding on the same bench was 13.69. same weights, same box. the gain came from the engine and from how many tokens you guess ahead, not from new hardware. the models got smaller for what they can do. the runtimes got a lot less wasteful. those two curves are meeting in the middle, and…
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-flash-next-GSQ-RSO-IQ3_S
Oct 31, 2026 · by @ItsmeAjayKV
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 8K | IQ3_S | strata | 86.5 tok/s | 1336.0 tok/s |
| 16K | IQ3_S | strata | 95.9 tok/s | 1707.0 tok/s |
| 32K | IQ3_S | strata | 80.9 tok/s | 1834.0 tok/s |
| 64K | IQ3_S | strata | 85.9 tok/s | 2093.0 tok/s |
Latest benchmark of Qwen3.8-flash-next-GSQ-RSO-IQ3_S with Strata on my 3090 I haven't updated Strata in a bit, the speed increase is insane. 8k: - tg: 86.5 tok/s - pp: 1336 tok/s 16k: - tg: 95.9 tok/s - pp: 1707 tok/s 32k: - tg: 80.9 tok/s - pp: 1834 tok/s 64k: - tg: 85.9 tok/s - pp: 2093 tok/s Have few more tasks to run with flash-next, improved speed will shorten the run time by a good amount hence allowing me to…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
RTX 3090 + 5060 Ti × Qwen3.8 27B
Oct 31, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | 6-bit | — | 55.0 tok/s | — |
| — | 2-bit Strata | — | 113.0 tok/s | — |
👀 Flash-Next hit 113 tok/s. Island still rough. Same voxel prompt on RTX 3090 + 5060 Ti: Qwen3.8 27B (6-bit) 55 tok/s, 123 min, every detail. Flash-Next 2-bit Strata done in 80 min, but skipped a lot. Speed isn't the whole creative story. via @0xironyAditya 👇 https://t.co/YVR4oNnnzn
Digest-only: confidence 0.55 below 0.60; incomplete methodology (engine+version, context length, hardware state, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 10 posts from the same window are listed here in short form, each with its source post.
- M5 Ultra × DeepSeek V4.1 — 241.1 tok/s decode (5 measurements) · @jundotkim
- Mac Studio M5 Ultra 256GB × Qwen3.8-27B — 1903.0 tok/s prefill (8 measurements) · @niklaslenz_ai
- H200 x4 × Deepseek-v4.1-flash-1017 — 150.0 tok/s decode (1 measurement) · @0xSero
- MI300X × Qwen3.8-Flash-Next — 33.5 tok/s decode (2 measurements) · @no_stp_on_snek
- RTX 3090 + RTX 5060 Ti × Qwen3.8 27B — 113.0 tok/s decode (2 measurements) · @0xironyAditya
- RTX 3060 × Qwen 3.8 Flash Next — 22.0 tok/s decode (1 measurement) · @jurlycat
- M3 Ultra × Qwen3.8 — 91.6 tok/s decode (1 measurement) · @QwenAIFans
- Unspecified platform × DeepSeek-V4.1-Flash — 10.0 tok/s decode (1 measurement) · @jun1228909
- Unspecified platform × MiniMax M78 — 39.0 tok/s decode (1 measurement) · @NFT_Chen
- Colibrì × Qwen3.8-Flash-Next — 45.7 tok/s decode (2 measurements) · @QwenAIFans
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
- @superalesha — Nov 1, 2026
- @niklaslenz_ai — Nov 1, 2026
- @QwenAIFans — Nov 1, 2026
- @MiaAI_lab — Nov 1, 2026
- @alizeh_ai — Nov 1, 2026
- @ivanfioravanti — Nov 1, 2026
- @EzoLabsInc — Nov 1, 2026
- @GuideboardLabs — Nov 1, 2026
- @murasametech — Nov 1, 2026
- @0xBakeer — Oct 31, 2026
- @ItsmeAjayKV — Oct 31, 2026
- @QwenAIFans — Oct 31, 2026
- @jundotkim — Oct 31, 2026
- @niklaslenz_ai — Oct 31, 2026
- @0xSero — Oct 31, 2026
- @no_stp_on_snek — Oct 31, 2026
- @0xironyAditya — Oct 31, 2026
- @jurlycat — Oct 31, 2026
- @QwenAIFans — Oct 31, 2026
- @jun1228909 — Oct 31, 2026
- @NFT_Chen — Oct 31, 2026
- @QwenAIFans — Oct 31, 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.