Cactus released Needle 3, an 8-29MB automation model matching DeepSeek V4 Flash.
One weight set, depths 2-20 layers, each its own model. Runs on Pi 5 at up to 4,000 tok/s. Not for chat, function calls only, returns empty if no tool fits.
Top-5 Best Value #LLM Models of the Day at UTC-16
| Model | #AAII | Price | | GLM 5.3 Flash | 42 | $0.09 | | DeepSeek V4.1 Flash | 40 | $0.14 | | DeepSeek V4 Flash 0731 | 34 | $0.04 | | GLM 5.3 | 45 | $0.98 | | GPT-5.6 Luna | 38 | $0.18 |
Pelican tests Deepseek 4.1 flash and ranks third in the world
Something suddenly occurred to me: would it be much faster if the front-end did not let the model write code and only let it make selections?
Then I tried it with the same requirement. Jev does not write code, only makes decisions, and the interface is directly built out of its own component library. In 0.7 seconds, the page is clickable. The same thing was thrown to DeepSeek V4.1 Flash, and it was still writing HTML honestly, 84…
You have reached the end of the archive
All of deepseek