Particle launches Particle-1 AI model for software engineering

Suhail Doshi, co-founder of Mixpanel, has launched Particle, a San Francisco-based AI startup. The company’s first product, Particle-1, is a post-trained version of DeepSeek V4.1 Flash designed to optimize software engineering workloads.

According to company evaluations, Particle-1 achieves similar accuracy to its base model while using an average of 45% fewer output tokens. Benchmark results across tasks such as Terminal-Bench 2.1 and ReactBench showed token reductions ranging from 28.4% to 52.2%, which translates to a reported cost reduction per task of between 34.4% and 51.6%.

The model was refined using reinforcement learning to reach correct answers more concisely. Particle-1 is currently available via API, with the company aiming to reduce task execution time and eventually expand into domains beyond software engineering. The startup is backed by Y Combinator, 8VC, and angel investors including Patrick Collison and Adam D’Angelo.