Particle releases a model trained to use fewer tokens on software tasks
Particle has released Particle-1 through its own application programming interface, focusing initially on software engineering. Built by further training DeepSeek V4.1 Flash, the model’s training encourages successful work while cutting the tokens used to reason and invoke tools. The company has published evaluations and examples spanning software and data engineering, while expansion into other fields remains a goal for later releases.
Artificial Intelligence··Evening
Software work is the first target
Particle, an AI model developer, released its first model through its own application programming interface on October 28. Particle-1 is based on DeepSeek V4.1 Flash, an existing language model, with further training directed at software engineering. The company’s reinforcement-learning reward favors solving a task while reducing the tokens spent on reasoning or on calling tools. Tokens are the units a language model processes or generates. Expansion into other fields is a goal for later releases.[1]
Training environments and task families
The company generated its training environments with Datamaker, its system for constructing such environments. Its evaluations place the model in terminal workflows and web applications built with React, as well as data-engineering and extended programming assignments. These include different kinds of work: compiling software in a terminal and fixing accessibility in a React application are among the examples. The model is offered through an API service, rather than a distribution of its weights for users to run independently.[1]
How the company examined completed work
Particle also published examples comparing successful attempts on the same task. They show shorter reasoning for a sequence of actions, fewer repair rounds and earlier stopping after completion. Some paired solutions differ in scope, including the alternatives explored and additional checks performed. The company’s average-cost calculations include unsuccessful attempts while excluding runs interrupted by infrastructure problems. A separate contract-review demonstration uses one legal task. The initial release and its main evaluation remain centered on software and data-engineering work.[1]
Related columns
For more information on this topic, you can read the related columns.
