Particle-1 is DeepSeek V4.1 Flash with a rank-64 LoRA adapter trained by reinforcement learning. The reward pays for solving the task and for using fewer tokens on reasoning and tool calls. Particle says it uses about 45% fewer tokens on average.
Particle's own numbers, across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1: 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower average cost per task than DeepSeek V4.1 Flash, with pass@1 inside the base model's confidence intervals. On one chart it shows Particle-1 at $0.2211 per task and a 72.57% pass rate, against $0.3368 and 71.15% for the base model. For that cost math it priced both models the same: $0.30 per million uncached input tokens and $1.20 per million output.
The post breaks down where the savings came from. About 90% of them are reasoning tokens on three of the four benchmarks: the model deliberates less, stops once a check passes, and writes shorter final confirmations. In one Terminal-Bench run the last two confirmation calls dropped from 11,445 to 1,765 tokens. Particle also flags a case where the savings came from a different (not identical) solution. This first release targets software engineering, through the Particle API.