Particle-1 Released: Optimized Based on DeepSeek, Task Success Rate Close to Original, Output Tokens Reduced by About 45%

Beating AI News Flash: Suhail Doshi, co-founder of product analytics platform Mixpanel and founder of AI image platform Playground, has started a new venture, founding Particle, and launched its first model, Particle-1. The model is post-trained with reinforcement learning based on DeepSeek V4.1 Flash, with a focus on reducing the tokens consumed by Coding Agents when executing tasks.

Four coding and data engineering evaluations released by Particle show that compared with the original DeepSeek, Particle-1 reduces output tokens by 28.4%–52.2%, averaging about 45%, while the task pass rate remains within the statistical confidence interval of the original version. Estimated at the same token unit price and a fixed cache hit rate, the cost per question drops by 34.4%–51.6%.

The saved tokens mainly come from more streamlined reasoning, fewer repeated attempts, and stopping promptly after task completion. In a Git deployment test, both models successfully completed the task, but Particle-1 output only 15,892 tokens, while the original version used 47,006.

Particle-1 has opened its API and is integrated into Vercel AI Gateway. The current performance and cost data mainly come from the team's own testing.