Unverified DeepSeek V4.1 Flash build claims 2000 tok/s prefill, 262k context
0xSero · x · 2026-10-03
User 0xSero announced a DeepSeek-V4.1-Flash-2-Sparks build (unverified by DeepSeek) with benchmark-style numbers: 2000 tok/s prefill, 29 tok/s prose decoding, 41 tok/s code decoding. The setup includes a 262k context window, a 2M-token KV cache pool, 8-way concurrency, and vision support.
Reported quality metrics: 92% top-token agreement and 0.07 KLD against the reference model, suggesting near-identical outputs. The build can be loaded via the pi tool. Authenticity remains unconfirmed.
More from Infra
- Tencent leaned on Japanese neocloud for AI compute; DeepSeek, Zhipu best positioned — teortaxesTex · 2026-11-19
- Open-source prefill engine for GLM-5.3-Flash beats vLLM by 3.2x on 4x B200 — CShorten30 · 2026-11-19
- Costco sees panic buying of generators after string of tech outages, 21 of 26 sold out — harris_edouard · 2026-11-19
- Starlink direct-to-cell goes live in Mongolia with free first month — XFreeze · 2026-11-19
- SpaceX unveils $50B, 6GW Arizona compute campus with 24 GWh grid-battery strategy — JOBhakdi · 2026-11-19
- CoreWeave corrects debt rate to 15.5% as analyst presses Blackwell-funded payoff plans — Svwang1 · 2026-11-19