Particle.ai launched Particle-1, a model post-trained on DeepSeek V4.1 Flash. It uses around 45% fewer tokens on average across tasks. Costs per task drop by 34.4% to 51.6% compared to the base model. Accuracy remains within the confidence intervals of DeepSeek V4.1 Flash. The first release targets software engineering tasks. It was trained using reinforcement learning with a custom reward function. The model uses fewer tokens for reasoning, tool calls, and verification. Particle-1 is available now via the Particle API. Early tests show token-efficient behavior may extend to legal and cybersecurity domains. Sign-up is required to access the API.
What changed
DeepSeek V4.1 Flash required more tokens and higher costs for the same tasks.
What it unlocks
Running software engineering tasks with fewer tokens and lower cost while maintaining accuracy.
- ~45% fewer tokens on average
- 28.4%–52.2% fewer output tokens
- 34.4%–51.6% lower cost per task
- 1.25% of model parameters (rank-64 LoRA)
What you need to act on it
- access to Particle API
- API key
Sources