Particle launches a coding model that uses 45% fewer tokens
Particle-1 post-trains DeepSeek V4.1 Flash for software agents, promising comparable task accuracy with lower cost and latency; the benchmarks are the company’s own.
Particle, a new AI lab founded by Playground and Mixpanel founder Suhail Doshi, has launched Particle-1, a coding-focused model designed to finish agent tasks with less reasoning and tool-call overhead. The model is available through Particle’s API.
Particle-1 is a reinforcement-learning post-train of DeepSeek V4.1 Flash. Particle says it used 28.4% to 52.2% fewer output tokens than the base model across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, while pass rates remained within the base model’s confidence intervals. Averaged across those tests, Particle reports a 72.6% pass rate and cost of $0.22 per task, versus 71.2% and $0.34 for DeepSeek V4.1 Flash.
The company says the savings come mainly from shorter reasoning, fewer repair loops and stopping after a task is verified. That matters for coding agents because output tokens drive both cost and latency, even when the underlying model price is unchanged. Particle lists the same per-token pricing as the DeepSeek base model.
Particle trained only a small LoRA adapter, equivalent to about 1.25% of the model’s parameters, rather than building a frontier model from scratch. The tradeoff is evidence: the benchmark suite and comparisons are self-reported, and there is no independent evaluation yet.
DeepSeek
