Particle-1

Particle-1 is Particle’s first model, designed to make capable AI cheaper and faster to use by solving tasks with fewer tokens. It is post-trained on DeepSeek V4.1 Flash using reinforcement learning that rewards successful task completion alongside lower reasoning and tool-call token use. On four software and data-engineering benchmarks - Terminal-Bench 2.1, ReactBench, ADE Bench, and DeepSWE 1.1 - it uses 28.4%-52.2% fewer output tokens and costs 34.4%-51.6% less per task than the base model, while pass rates remain within the base model’s confidence intervals. Its accuracy is also close to the base model on general benchmarks, including GPQA Diamond, Tau Airlines, MMLU, and BabyVision. The launch focuses on software engineering, with other domains planned.

The token savings mainly come from shorter deliberation, less repeated checking after a task is complete, fewer implementation-repair cycles, and more focused exploration. Reasoning accounts for about 90% of savings on three benchmarks; tool-call text also contributes substantially on DeepSWE. Example comparisons show Particle-1 reaching implementations earlier and using fewer calls in some cases, though savings can also reflect less extensive testing or a different solution. Training used synthetic coding and data-engineering environments with verifiable outcomes, in successive reinforcement-learning stages; coding-stage gains also appeared on data tasks, suggesting some generalization. The team reports early signs of efficiency carrying into legal and cybersecurity work, but says broader releases will follow.