Particle launches particle-1, an AI model using ~45% fewer tokens with performance parity. Y Combinator's Garry Tan calls it primed for the smartest model at the lowest cost per intelligence

Particle, a new company, has launched particle-1, its first model in a developer-focused series, trained for token efficiency. The model is based on DeepSeek V4.1 Flash 1017 and, per the author, completes end-to-end tasks noticeably faster. The company aims for continuous post-training, shipping model improvements as fast as software. Citing collapsing gaps between frontier and open-weight models, plus new chips and energy sources, the founders argue intelligence costs should approach zero and say they will work across inference, models, hardware, and energy. The company is hiring and thanked backers and advisors.

✨ AI summary of the linked source — not editorial content.

Explore the topic

• Launch
• AI