DeepSeek V4.1-Flash: open MIT weights beat Sol at agentic coding
552B MoE, 1M context, self-host or hit the API at $0.15/M — it edges Opus 5 too. Plus SWE-2 lands on Devin and DeepSeek retires V4 Pro.
Copy markdown
It edges Sol and Opus 5 where you code
On DeepSWE 1.1 it posts 74.2, past GPT-5.6 Sol (73.0) and Claude Opus 5 (74.0), and it tops Terminal-Bench 2.1 (90.6), AutomationBench (54.8) and Agent's Last Exam (31.8). For agent loops, an open model now trades blows with the frontier.
$0.15 in, and cache hits are almost free
Off-peak API pricing is $0.15/M input and $0.60/M output; a cache hit drops input to $0.003/M. That makes long, tool-heavy agent runs cheap enough to leave running.
1M context, 384K output, a KV cache that fits
FP4 quantization plus cross-layer attention shrink the KV cache to ~890 bytes per token — roughly 8x less memory than V4-Flash — so the 1M-token window and 384K-token outputs are actually servable on modest hardware.
MIT weights on Hugging Face — ship it commercially
The full weights are MIT-licensed and downloadable now, and the API speaks the OpenAI Responses/Codex format. Self-host it, or drop it into an existing OpenAI-shaped client with a base-URL swap.
The catch: long-horizon still belongs to Opus
On Terminal-Bench 4.0 it scores 31.2 versus Claude Opus 5's 51.8, and it trails on repo-scale work. Route the cheap Flash for bounded tasks and keep a frontier model for the multi-hour jobs.
Heads-up: V4 Pro retires Sept 14
DeepSeek is pulling V4 Pro on September 14, calling V4.1-Flash a full replacement. If you pin the V4 Pro endpoint anywhere, migrate this week before those requests start failing.
Elsewhere: Cognition's SWE-2 lands on Devin
SWE-2 post-trains Moonshot's Kimi K3 and hits 92.8% on Terminal-Bench 2.1 at roughly 64% less cost than Fable 5.1. It's live in Devin Desktop and CLI now, with trained reasoning-effort levels to dial cost against quality.