SITUATION EXPLAINED: DeepSeek V4.1 Flash beats its own Pro model from 3 weeks ago.
552B mixture of experts, 8B active for input and 16B for output. MIT licensed, 1M context • Beats V4 Pro on every benchmark except GPQA Diamond and Humanity's Last Exam • $0.15 input, $0.60
DeepSeek-V4.1-Flash on four DGX Sparks, single stream, live 🐋
The predecessor (V4 Flash + DSpark) peaked around 50 to 70 tok/s on code. V4.1 is 552B total, 16B active on the way out, 8B active on prefill, and about half the checkpoint is n-gram memory tables that we keep on
This campfire cost $0.10.
One prompt, 41 minutes, 1,398 lines. DeepSeek v4.1 Flash — Three.js, TSL, WebGPU, procedural fire and sound. The agent checked its own work: headless Chrome screenshots caught 3 rendering bugs, CDP confirmed the audio graph was actually running. 60
DeepSeek V4.1 Flash (High) vs Kimi K3 (Low)
You have reached the end of the archive
All of deepseek