DeepSeek V4.1-Flash is live on Nito.
Run it inside Claude Code or Codex, or in the Nito web app, fully incognito. By DeepSeek's numbers it tops Opus 5 and GPT-5.6 on all four agentic benchmarks, with 8B active parameters on input and a KV cache a quarter the size.
Told deepseek v4.1 flash to generate a pagoda garden and this is what came out.
Other than the trees being messed up I think this is really damn good.
Deepseek V4.1 Flash has DESTROYED Kimi K3 and GPT-5.6 Sol at coding
You won't find a better model for everyday use it runs at 300–400 t/s and is 15 times cheaper than the competition plus, it is both multimodal and efficient and only slightly worse than the best models in the
You have reached the end of the archive
All of deepseek