DeepSeek-V4.1-Flash lands: MIT weights, 1M context, cheaper Flash rates

DeepSeek released V4.1-Flash, a 552B MoE that activates 8B/16B params, under MIT with 1M context and cut Flash API rates; it has since dropped a plan to route all V4-Pro traffic to it from Sept 14.

Key points

• API id deepseek-flash; weights at deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face, MIT licence. 552B MoE, 20-layer encoder + 20-layer decoder; 8B active on prefill, 16B on decode.
• 1M context, 384K max output on the API, native image+text input. 45T-token pretrain; FP4 KV cache that DeepSeek says needs 1/4 the HBM and 1/8 the SSD of V4-Flash.
• Off-peak per 1M tokens: $0.003 cache hit, $0.15 miss, $0.60 output; peak doubles. Vs V4-Flash-0731 that is -57%, -32%, -9% (our arithmetic); live since 04:00 UTC Sept 10.
• DeepSeek's launch post said every deepseek-v4-pro call would go to V4.1-Flash at Flash rates from 04:00 UTC Sept 14; its pricing page now says V4-Pro stays on the API after Sept 14, billed as before. V4-Flash and V4-Flash-Vision-Exp are already retired.
• DeepSeek's own table: DeepSWE v1.1 74.2 (V4-Pro 62.7, Opus-5.0 74.0), Terminal-Bench 2.1 90.6 (V4-Pro 87.9), CyberGym 88.1. It trails Opus-5.0 on Terminal-Bench 3.0 (30.0 vs 43.3) and 4.0 (31.2 vs 51.8), ProgramBench (20.3 vs 37.0), NL2Repo (64.0 vs 75.3) and HLE (36.8 vs 56.3), and trails V4-Pro on GPQA Diamond (90.9 vs 92.4).
• Live on the DeepSeek API via both its OpenAI-format and Anthropic-format base URLs; Novita is listed as an inference provider on the HF page. The model repo has a minimal inference folder and a Python prompt-encoding reference; DeepSeek also released deepseek-recipe, separate Rust libraries with Python bindings, for production prompt encoding. vLLM and SGLang have posted deployment guides (vLLM, SGLang) that run it from special Docker builds rather than a regular pip release.

Beyond The Next Web's report, the specs and benchmark scores here are from DeepSeek's model card, the API name, limits, prices and the V4-Pro reversal from its pricing page (the V4-Flash-0731 prices from that page as archived on Sept 9), and the KV-cache savings, retirements and quote from its launch post.

“We're passing the savings on to you.” — DeepSeek, V4.1-Flash launch post

Why it matters

For a small studio the practical story is the input side: an agent loop that re-reads a repo on every tool call now pays $0.003 per million cached tokens off-peak, and the 8B-active prefill was built for exactly that pattern. Coding results are uneven: it matches Opus-5.0 on DeepSWE but trails it by 13 to 21 points on Terminal-Bench 3.0 and 4.0 and ProgramBench, and every score here is DeepSeek's own, so run your own tasks before swapping. Anyone still calling deepseek-v4-flash is already being served by the new model without a code change, so watch for behaviour drift.

This is an AI-written summary of the reporting credited above and the other pages linked in the text, read and edited by Nicholas before publishing. The facts belong to those sources and the quote to DeepSeek; the wording is ours. Read the original.