Suhail launches Particle with token-efficient particle-1 model

Suhail announced a new company called Particle on October 28, describing it as a venture focused on building efficient AI models for developers. Its first model, particle-1, is post-trained on DeepSeek V4.1 Flash and is designed to use fewer tokens per task rather than add new capabilities. According to Particle's own writeup, the model uses roughly 45% fewer tokens on average across a range of tasks while keeping accuracy close to its base model, and Suhail said completing tasks end-to-end is noticeably faster as a result.

Particle's blog post detailed the training approach: reinforcement learning in internally built coding and data-engineering environments, using techniques including Group Sequence Policy Optimization and a LoRA adapter covering all layers. Across four benchmarks covering terminal workflows, React app building, data engineering, and longer software engineering tasks, the company says particle-1 cut output tokens by 28.4% to 52.2% and cost per task by 34.4% to 51.6% compared with DeepSeek V4.1 Flash, with pass rates staying within the base model's confidence intervals. The company also reported early, smaller-scale signs that the token savings carry over to tasks outside coding, including a legal contract-review example.

Particle-1 is available now through the Particle API, and Suhail separately announced a partnership with OpenRouter to offer inference for the model.

Why it matters

If the efficiency gains hold up independently, cheaper and faster agent runs could meaningfully cut the operating cost of coding agents built on open-weight models.

Source: Suhail on X