DeepSeek V4 Flash: MIT coding model beats its bigger V4-Pro sibling
The 284B/13B-active MoE is topping Hacker News as builders clock sub-$0.20 agent runs — while Oracle bars all AI-written code from OpenJDK.
Copy markdown
It beats the model twice its size
V4 Flash — a 284B-parameter MoE with just 13B active per token — tops DeepSeek's larger V4-Pro (Preview) on all nine published agentic and coding benchmarks; Terminal-Bench 2.1 jumped from 61.8 to 82.7. The weights are MIT-licensed with a 1M-token context.
Agent loops for pennies
DeepSeek's API lists $0.14 per million input tokens and $0.28 output (cache hits drop input to $0.0028), and OpenRouter already fans it across 23 providers from as low as $0.09/$0.18 — cheap enough to leave a coding agent grinding all weekend.
Or run it on your own box
The ungated MIT weights are on Hugging Face and need roughly 110GB of combined VRAM/RAM for a single node; DeepSeek's DSpark speculative decoding claims 60–85% faster per-user generation. It suggests temperature 1.0, top_p 0.95 for agent workflows.
Where the open models can't go
Oracle's interim OpenJDK policy — now blowing up on Hacker News — bars any code, text, or images 'generated in part or in full' by LLMs or diffusion models from contributions, citing IP risk and 'plausible-looking but incorrect code,' even as Larry Ellison touts AI writing Oracle's own software.
Elsewhere: Kimi K3 lands in GitHub Copilot
GitHub has resumed rolling out Moonshot's Kimi K3 inside Copilot, billed at provider list pricing under usage-based billing — another open-weights contender now one dropdown away in your IDE.