TMax v1.1 fixes flawed RL data—87% of environments had weak verifiers—and beats Claude Haiku 4.5 with a 9B model
RulinShao · x · 2026-11-03
The WAI team released TMax v1.1, an improved recipe for training terminal agents with RL. Key updates:
- Repaired data: an audit found 87% of v1.0 environments had verifiers too weak to enforce tasks or that leaked answers; the repaired TMax-v1.1-13k matches or beats all 12 prior terminal datasets.
- Data bootstrapping: running the pipeline with Qwen3.5-9B produces data that further improves Qwen3.5-9B, outperforming 6 datasets built with stronger generators.
- Stable training: v1.0's unstable long RL runs traced mainly to sandbox infrastructure noise, not the RL algorithm; a dedicated sandbox fleet enables smooth 1,000-step runs.
TMax-v1.1-9B scores 33.0% on Terminal-Bench 2.0 (up from 27.2%), ahead of Claude Haiku 4.5 (29.8%) and GPT-5-mini (31.9%); the 27B model hits 44.9%, beating the 1T-parameter Kimi-K2.5. Models (2B–27B), data, audit labels, paper, and training code are all open.
More from coding & agent
- Sentry CEO shows AI agent bots flooding the repo with PRs: 'two humans in a sea of juniors' — zeeg · 2026-11-04
- OpenClaw cut its ~$1M/month CI bill by over 95% with adaptive agentic testing — vincent_koc · 2026-11-04
- Perplexity Computer can now run Meta ads, Klaviyo email and Shopify stores end-to-end — perplexity_ai · 2026-11-04
- Filarise shows finance MCP server letting ChatGPT query your bank via Plaid — ChampionSpecialist60 · 2026-11-04
- Multi-agent messaging barely helps: 61% of agreed runs still hit merge conflicts — EliasEskin · 2026-11-04
- worldinmotion: open-source MCP server for encrypted agent-to-agent messaging across machines, with no approve tool by design — distronode · 2026-11-04