Particle-1 post-trained on DeepSeek V4.1 Flash cuts output tokens ~45%
heyneighbor · x · 2026-11-04
Particle launched Particle-1, post-trained on DeepSeek V4.1 Flash using internal RL environments that reward solving tasks with fewer reasoning and tool-call tokens. It uses 45% fewer output tokens on average with comparable accuracy.
- First release targets software engineering; other domains planned
- On Terminal-Bench 2.1, DeepSWE 1.1, ReactBench and ADE Bench, costs drop sharply: DeepSWE at $0.22/task and 59,580 output tokens vs 83,246, Terminal-Bench 3x cheaper
- Standard inference comes with zero data retention
- Available in OpenClaw v2026.10.10
More from coding & agent
- Wanaka Studio 1.0 launches as an AI-native game engine claiming 5x speed and 82% lower agent cost — bdsqlsz · 2026-11-04
- Intology's Locus autonomously adapts Boltz-2 and frontier scientific models to proprietary data, claiming SOTA-beating molecular design — anshulkundaje · 2026-11-04
- Traversal raises $120M at $1B valuation to keep AI-generated code running in production — kimmonismus · 2026-11-04
- OpenAI merges ChatGPT's Chat and Work tabs into one interface — emmanuelvivier · 2026-11-04
- Dotey claims Astra codes worse than Opus 5.5,vczh fires back: tell it what code you want — dotey · 2026-11-04
- New 10-Week Course Ships One Real Product Using AI Coding Tools — Al_Grigor · 2026-11-04