DeepSeek V4.1 Flash: MIT open weights at $0.15/M, 1M context
A 552B MoE to self-host or call for pennies, coding-strong — while Cognition's Devin-only SWE-2 shipped the same day to benchmark-gaming pushback.
Copy markdown
Open weights, MIT, on Hugging Face
DeepSeek released V4.1 Flash under a permissive MIT license: a 552B-parameter mixture-of-experts that activates only 8B per token, so you can self-host the weights or call the hosted `deepseek-flash` endpoint. The weights and quantized builds are already up on Hugging Face.
Pennies per million tokens
Off-peak input is $0.15/M and output $0.60/M, with cache hits at $0.003/M; peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays) double that. That puts near-frontier coding at a fraction of Sol, Fable, or Astra pricing.
1M context, and it actually codes
The model takes up to 1M tokens in and emits up to 384K out, and posts 74.2 on DeepSWE v1.1 (real bug fixes in real repos) and 90.6 on Terminal-Bench 2.1. Build this weekend: a whole-repo refactor agent that reads your entire codebase in one call for cents.
Heads-up: v4-pro reroutes Sept 14
From September 14, DeepSeek routes every `deepseek-v4-pro` request to V4.1 Flash and bills it at Flash prices. If your code pins v4-pro, expect a different model behind the same name — and a smaller bill — starting Monday.
SWE-2 shipped the same day — but Devin-only
Cognition's SWE-2 claims 92.8 on Terminal-Bench 2.1 at 64% less cost, but it runs only inside Devin (Desktop, CLI, Web) — no API, no open weights. It also drops to 27.3 on the harder Terminal-Bench 4, drawing benchmark-optimization skepticism the open, callable DeepSeek sidesteps.