Devlog #3 Nanoharness I'm not sure if I calculate TPS correctly
But deepseek V4.1 flash seems incredibly fast despite only 100t/s Also I think I need to work on cache hit, or check the calculation again
Atomic Chat ran a test: feed DeepSeek-V4.1-Flash a prompt
Output a single local HTML file, and ask for a ray-traced FPS with weapon mechanics and enemy waves. It worked. No asset pipeline. No UI framework. One pass. The model is a 552B MoE with a Causal-Encoder-Decoder and
This Anthropic report release was pretty wild
Especially the section where they mentioned how Kimi and DeepSeek routes API calls to Opus model sometimes. I have some thoughts, and showing you encrypted reasoning traces that Anthropic says Chinese labs have been stealing.
You can now run DeepSeek-V4.1-Flash on your iPhone using with a 2 GB memory budget.
You have reached the end of the archive
All of deepseek