Training a Chinese model to "answer more briefly"
Suhail Doshi, who founded analytics tool maker Mixpanel, announced on X on Oct. 28 (Oct. 29 Japan time) that he had launched a new company, Particle, and released its first model, "Particle-1." Doshi also founded Playground, a U.S. image generation AI company.
Particle-1 is built on "V4.1 Flash," a small model from China's DeepSeek, and further trained for software development and data processing tasks. The aim is to reduce the tokens used for reasoning and tool calls without lowering accuracy. The longer the tasks handed to an AI, the more output tokens pile up, so it lowers that portion of the cost. The weights have not been released; the model is offered through an API and is available via OpenRouter and Vercel.
About 45% less output, accuracy on par
Particle tested the model on four development benchmarks, including terminal operations and code changes. Output tokens fell by about 45% on average (28% to 52% depending on the benchmark), and the cost per task fell by 34% to 52%, the company said. Accuracy was on par with the base model within the margin of error.
| Model | Cost per task ($) | Output tokens | Accuracy |
|---|---|---|---|
| Particle-1 | 0.22 | 59,580 | 72.6% |
| DeepSeek V4.1 Flash | 0.34 | 83,246 | 71.2% |
| GLM 5.3 Flash | 0.65 | 98,155 | 66.8% |
Some verification steps skipped
All figures are from Particle's own evaluation and have not yet been reproduced by third parties. An explainer article noted that part of the savings came from reducing verification steps and stopping tasks early. The costs are calculated on the assumption that 95% of input can be reused from cache. Pricing figures differ between Particle's announcement and the OpenRouter listing.
