BAAI Open-Sources AREX-2: 27B Long-Horizon Self-Improving Agent
On September 30, BAAI released and open-sourced AREX-2, a 27B long-horizon agent built on the Qwen3.8 27B base with a 262,144-token context window under Apache 2.0. Weights are live on Hugging Face with GGUF versions, and the model scores 92.2 on GAIA and 81.8 on MLE-bench Lite.
Confirmed
- Open-source details: Apache 2.0 license for commercial use; weights on Hugging Face with GGUF builds for local deployment (@Skyline34rGt, @AdinaYakup)
- Approach: per BAAI, self-improvement is decomposed into reflection (producing better solutions than the current one) and long-horizon execution (not degrading over many rounds), assumed domain-agnostic and learnable in feedback-verifiable domains (@BAAI)
- Training data: synthetic long-horizon improvement trajectories from ML and algorithmic coding tasks, learned via execution feedback (@BAAI, @akhaliq)
- Benchmarks: 92.2 on GAIA (@BAAI), 81.8 on MLE-bench Lite (@akhaliq)
- A key property is that more test-time rounds yield better solutions (@akhaliq)
Why it matters
- Self-improvement and long-horizon execution are key bottlenecks for agents; AREX-2 offers a systematic training recipe in verifiable domains
- 262K context plus a commercial-friendly Apache 2.0 license makes it practical for developers deploying long-horizon agents locally
2026-09-30 ~ 2026-10-02 · 5 related posts
Primary sources
- BAAI Releases 27B Long-Horizon Agent Model AREX-2 with 262K Context — Skyline34rGt · 2026-09-30
- BAAI open-sources AREX-2, a 27B long-horizon agent model with 262K context under Apache 2.0 — AdinaYakup · 2026-09-30
- [source] BAAI's AREX-2 Trains Self-Improving Agents, Hits 92.2 on GAIA and 81.8 on MLE-bench Lite — BAAI · 2026-10-01
- AREX-2: 27B reflective agent scores 81.8 on MLE-bench Lite via long-horizon self-improvement — _akhaliq · 2026-10-02
1 near-duplicate retellings: AdinaYakup