X local-model bench digest — Sep 14, 2026
As of Sep 14, 2026 — 12 posts collected from X. 25 measured runs catalogued, 12 kept to the digest only. Latest source post Sep 14, 2026 07:13 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 × DeepSeek-V4.1-Flash
Sep 14, 2026 · by @MiaAI_lab
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 1M | — | — | 45.0 tok/s | — |
| 1M | — | — | 73.0 tok/s | — |
| 1M | — | — | 134.0 tok/s | — |
Run DeepSeek v4.1 Flash on your 4x DGX Sparks ✨ - 8,000,000 KV cache pool - That's correct - it's 8 million (!) - 1M context & 8 concurrent sessions by default - 1M context needle passed - Official weights, no quantitation! This means you can now run 8 concurrent sessions, each with a 1M context window and its own KV pool! These are confirmed, conservative settings for stability. KV cache pool could probably be…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
Unspecified platform × Qwen
Sep 14, 2026 · by @coinjoined · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 0.7 tok/s | — |
be me in 2026 @sama and @DarioAmodei say open-weight AI needs more guardrails everyone argues about AI policy not me I've seen Johnny Mnemonic I know how this ends start downloading every open-weight model I can find Qwen Llama Mistral DeepSeek some 900GB torrent called FINAL_UNCENSORED_REAL_v7_FIXED no idea what it is seed it anyway buy 47 SSDs bury three in the woods give one to my lawyer freeze one in a block of…
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 4.1 Flash
Sep 14, 2026 · by @melvindvivas · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 325.0 tok/s | — |
This is how fast DeepSeek 4.1 Flash is using their official API 325 tok/s https://t.co/aGUQ7cLhed
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
Sep 14, 2026 · by @NFT_Chen
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | EXL3 2.9bpw | vllm | 32.0 tok/s | — |
| — | EXL3 2.9bpw | vllm | 42.0 tok/s | — |
| 8K | EXL3 2.9bpw | vllm | — | 1000.0 tok/s |
| 256K | EXL3 2.9bpw | vllm | — | 872.0 tok/s |
🔥推荐：仅两台DGX Spark，把当前最能打的开源模型DeepSeek V4.1 Flash跑起来了，且性能只升不降！ DeepSeek-V4.1-Flash 不再是 3/4 机专属。 性能亮点： 🔹散文单流 ≈32 tok/s 🔹双并发合计 ≈42 tok/s 🔹8k–128k prefill ≈1000 tok/s 🔹256k prefill ≈872 tok/s 🔹默认 600k 上下文 / 775k KV 🔹550k 长 prefill 压测通过 🔹EXL3 2.9bpw + Engram 落盘，原生大权重直接塞进 2 机 🔹DSpark 投机解码、工具调用、视觉都在 保守默认，先求稳，再谈猛。 GitHub仓库：https://t.co/9wkvKphEol #DeepSeek #DeepSeekV41Flash #DGXSpark #LocalLLM #vLLM #EXL3 #开源大模型 #本地部署 #梁文峰
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
RTX 5090 × Qwen3.8-27B
Sep 14, 2026 · by @TeksEdge · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 256K | Q4_K_S | — | 75.0 tok/s | — |
Okay, this is interesting! A new Qwen3.8-27B post-train added a "transmission". 🏎️ "Qwen3.8-27B Twin Turbo" is a community-tuned multimodal GGUF that lets you switch between a bunch of stuff like ... 🧠 5 reasoning modes ⚡ 5 non-thinking/instruct modes 👁️ Vision 📚 256K context 🚀 MTP speculative decoding …all inside ONE locally running 27B model. But they aren't trying to simply optimize tps; they are attempting to…
Digest-only: confidence 0.35 below 0.60; incomplete methodology (engine+version, hardware state, measurement method).
DGX Spark x4 + RTX 5090 × DeepSeek-V4.1-Flash
Sep 14, 2026 · by @dangerm00se
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 1M | — | ds41rt afd | — | 2000.0 tok/s |
| 1M | — | ds41rt afd | 79.0 tok/s | — |
DeepSeek-V4.1-Flash (769B w/ Engram) at full 1M context on 4× DGX Spark + one RTX 5090, running @wrldsuksgo2mars DS41RT AFD engine with the coordinator ported to the 32 GB card. 2,000 tok/s prefill, 73–85 tok/s decode, 7.79M-token KV pool, 8 concurrent 1M requests. https://t.co/yPuCYb3gPs
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state, measurement method).
AMD Radeon Pro V620 x4 × Qwen3.8-Next-Flash
Sep 14, 2026 · by @jurlycat
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 128K | W4A16 | — | — | 1300.0 tok/s |
| 128K | W4A16 | — | 65.0 tok/s | — |
You don’t need a $20,000 server to run large AI models locally. One builder assembled a 128GB VRAM inference machine for roughly $3,000 using mostly old data-center hardware: • 4× used AMD Radeon Pro V620 GPUs — 32GB GDDR6 each • AMD EPYC 7452 • 256GB DDR4 RDIMM • Huananzhi D12D motherboard • 1600W PSU On Qwen3.8-Next-Flash AutoRound W4A16, the builder reports 1,300 tok/s prefill and 60–70 tok/s generation at 128K+…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
Unspecified platform × Qwen
Sep 13, 2026 · by @Italianclownz · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | MXFP4 | — | 236.0 tok/s | — |
There was a post by Alex K Miller that discussed using Radiance + Qwen 3.8 27B MXFP4 to get good Prefill, TTFT, and Decode. I replicated it on Linux and got around the same as his article. Then I tried a derived version of @OrcaRouter 's posted on hugging face https://t.co/NAyuLmP7qX and got even higher numbers. 236 tok/s decode speed on reasoning. Original Post by Alex K Miller https://t.co/xE6JYiLJnb Thanks to…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, context length, hardware state, measurement method).
DGX Spark x1 or x2 × DeepSeek-V4.1-Flash
Sep 13, 2026 · by @0x0SojalSec
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 600K | EXL3 | — | 32.0 tok/s | — |
| 600K | EXL3 | — | 42.0 tok/s | — |
| 600K | EXL3 | — | — | 1000.0 tok/s |
You can now run locally one of the strongest models DeepSeek v4.1 Flash on just one or two DGX Sparks. - 600k context/ 775k KV cache - 32 tok/s single stream, - 42 tok/s with 2 concurrent - 1000 tok/s prefill. - EXL3 quantization One of the best models you can actually run today not need a data center - https://t.co/UWg14Qeu0X
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
Unspecified platform × qwen3.8
Sep 13, 2026 · by @AlexanderLong · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 200.0 tok/s | — |
My position is typically to focus on getting things to work rather than making noise about what needs to be done, but enough people have messaged me in last 24 hours that felt I should write some brief thoughts. The tl;dr is I completely agree with the sentiment but think no-one aside from Pluralis is earnestly moving towards a realistic solution. The reality is without some mechanism to pool resources for the…
Digest-only: confidence 0.35 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Jovian Judgement R37/TP4/DCP1 × DeepSeek V4.1 Flash
Sep 13, 2026 · by @0x0SojalSec
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 32768 | — | — | — | 21400.0 tok/s |
| — | — | — | 280.0 tok/s | — |
| — | — | — | 500.0 tok/s | — |
| — | — | — | 417.0 tok/s | — |
You can Run Locally DeepSeek V4.1 Flash It beats GPT-5.6 Sol, Claude Opus 5 & can run with 4-GPU at your home locally No NVLink. - Jovian Judgement R37/TP4/DCP1: - 21,400 tok/s uncached 32K prefill - 280 tok/s C1 decode (budget 50) - 500 tok/s single-stream decode - 417 tok/s sieve - +44% prefill vs R36 - DSpark K7 + Engram (RAM or SSD) - 552B MoE, vision, 1M context Engram in pinned RAM or disk, DSpark K7. -…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state, measurement method).
DGX Spark x3 × DeepSeek-V4.1-Flash
Sep 13, 2026 · by @jakeharrisdev
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | jspark3 v2.0.0 (Tempo) | 71.5 tok/s | — |
| — | — | jspark3 v2.0.0 (Tempo) | 33.5 tok/s | — |
JSPARK3 v2.0.0 (Tempo) is live. JSPARK v2.0.0 is running DeepSeek-V4.1 Flash, meant to run locally on three DGX Sparks. Single-stream generation: ⚡ 67–76 tok/s code ⚡ 33–34 tok/s prose It's my new daily driver. Recipe, benchmarks, and setup guide are all here: https://t.co/sIXIcTDLnV
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (quantization, context length, hardware state, measurement method).
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
- @MiaAI_lab — Sep 14, 2026
- @coinjoined — Sep 14, 2026
- @melvindvivas — Sep 14, 2026
- @NFT_Chen — Sep 14, 2026
- @TeksEdge — Sep 14, 2026
- @dangerm00se — Sep 14, 2026
- @jurlycat — Sep 14, 2026
- @Italianclownz — Sep 13, 2026
- @0x0SojalSec — Sep 13, 2026
- @AlexanderLong — Sep 13, 2026
- @0x0SojalSec — Sep 13, 2026
- @jakeharrisdev — Sep 13, 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.