Back home

Particle-1

x+web 0.80
Visit source

Intro

Particle-1 is a token-efficient agentic coding model from Particle, a new AI inference company announced by Suhail (founder of Playground) on October 28, 2026. It is post-trained on DeepSeek V4.1 Flash and keeps accuracy at parity while using about 45% fewer tokens on average, so tasks cost less and finish sooner.

Announcement: Introducing Particle-1 · Launch post: Suhail on X · API listing: OpenRouter

Explain

Key points

  1. What Particle is: a new inference layer for open-weight models (fast, low-cost, reliable). Its API also serves DeepSeek V4.1 Flash and GLM 5.3 Flash. Particle-1 is its first in-house model.
  2. Token efficiency: across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, Particle-1 uses 28.4% to 52.2% fewer output tokens and has 34.4% to 51.6% lower average cost per task than base DeepSeek V4.1 Flash, with pass@1 within the base model's confidence intervals.
  3. Headline numbers (Particle's chart): cost per task $0.2211 vs $0.3368 (DeepSeek V4.1 Flash) and $0.6523 (GLM 5.3 Flash); mean output tokens 59,580 vs 83,246 and 98,155; task pass rate 72.57% vs 71.15% and 66.82%.
  4. General skills kept: GPQA Diamond 90.40% (base 90.91%), Tau Airlines 86.00% (88.00%), MMLU 86.08% (85.98%).
  5. How it was trained: reinforcement learning in internal coding and data-engineering environments (Harbor format, built by its synthetic data system Datamaker, run through the Terminus 2 agent), staged Cascade RL, the open-source Miles RL framework, a rank-64 LoRA over all experts, and GSPO for policy updates.
  6. Scope: this first release focuses on software engineering; Particle plans other domains later.

Usage

How to use it

  • Particle API: sign up at particle.ai.
  • OpenRouter: model id particle-ai/particle-1.
  • Vercel AI Gateway: model id particle/particle-1.

Good fits

  • Debugging, planning code changes and breaking down multi-step engineering problems
  • Terminal and data-engineering agents (for example dbt projects)
  • High-volume coding agents where token spend is the main cost

Tip

If you already run DeepSeek V4.1 Flash for coding agents, swap in Particle-1 on the same tasks and compare cost and wall-clock time per task.