by BAAI
27B open-weights agent model from BAAI that iteratively proposes, tests and refines solutions on coding and ML tasks.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
AREX-2 is a 27B-parameter dense agent model from the Beijing Academy of Artificial Intelligence (BAAI), built on a Qwen3.8-compatible multimodal base. It is trained to improve a solution over multiple test-time rounds: propose, measure, reflect and revise, using feedback such as scores, logs, errors and timings.
Training uses machine-learning and algorithmic-programming tasks with verifiable feedback, combined with the earlier AREX deep-research data. According to the model card, the self-improvement behaviour transfers to deep research without new search trajectories. The card reports results on Frontis-CS-style coding and MLE-Lite benchmarks against closed and open-weight models, including GPT-5.6 Sol and Kimi-K3; these are vendor-reported figures following the paper's protocols. The context length is 262,144 tokens.
Weights are published on Hugging Face under Apache-2.0 in safetensors format for use with the transformers library (roughly a 55 GB download, per a creator review). It targets researchers and developers running long-horizon agentic coding, ML-engineering and deep-research workflows locally or on their own infrastructure.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
A 27B Qwen-based model trained to iteratively propose, test and refine its own solutions on verifiable ML and algorithmic tasks. It matches much larger models, including Kimi K3 and GPT 5.6 Sol on some benchmarks, and is a 55 GB download.
Closest neighbours by meaning, not just by category.
Model
35B Mixture-of-Experts open-weights model from InternScience for long-horizon agentic search, engineering and tool use.
Model · Meituan
Meituan's open-weights 1.6T-parameter mixture-of-experts LLM (48B active), aimed at agentic coding and reasoning tasks.
Model · Nex AGI
Open agentic model family from Nex AGI for computer use, coding and long-horizon tool tasks, in several sizes.
Model · Alibaba
Alibaba's open-weights 2.4T-parameter MoE language model (95B active) from the Qwen3.8 family, with a hosted Max version.
Model · IQuest Lab
320B-parameter MoE model (15B active) from IQuest for agentic coding and tool use, with 512K context.
Model
Open-weights 35B-A3B agentic model post-trained from Qwen3.6-35B-A3B for search, coding and scientific research.