BAAI's AREX-2 trains long-horizon reflection on a 27B base: benchmark scopes differ, weights posted but license unstated
According to Machine Heart (Jiqizhixin), the AREX research group trained AREX-2 on the Qwen3.8-27B base, reporting MLE-bench Lite 81.8, Frontier-CS 70.7, BrowseComp 84.0 and HLE 52.6, with links to a paper and a Hugging Face weights collection; the report does not state the license, context length or a hosted API.
What changed
Machine Heart reports that the AREX research group, with researchers from the Beijing Academy of Artificial Intelligence, the University of Science and Technology of China, Renmin University of China and other institutions, released AREX-2, built on the Qwen3.8-27B base and trained on "long-horizon reflection" data constructed around machine learning and algorithmic programming scenarios. The figures and their respective comparison scopes are: MLE-bench Lite 81.8, above every comparison model listed in this report (including closed frontier models, with GPT-5.6 Sol and Kimi-K3 both at 72.7); Frontier-CS 70.7, ahead of all open-weight models in the table that report a score for this item; BrowseComp 84.0 and HLE 52.6, which the report says are the highest scores among models with no more than 40B parameters. The report also says BrowseComp 84.0 exceeds the 1T-parameter Kimi-K2.6 (83.2) and the 1.6T-parameter DeepSeek-V4-Pro (83.4). On training, the report says teacher agents first autonomously build tasks, scoring mechanisms and runtime environments, then loop under time and turn budgets through "submit a plan, get feedback, reflect and improve", fully preserving reasoning, failure and adjustment trajectories; the deep research evaluation reuses AREX v1 data and only adds long-horizon reflection trajectories, to test cross-domain transfer. The article ends with links to an arXiv paper, a GitHub project page and a Hugging Face weights collection, but does not state the license type, commercial terms, context length, inference overhead or any hosted API.
What it means for developers
The direct implication for developers is that this looks more like a downloadable open-weight candidate than a callable interface service. The report gives only a weights collection address, with no mention of vendor hosting, rate limits or pricing, so access requires self-deployment or waiting for third-party hosting, and VRAM, throughput and inference cost must be assessed by the user. The benchmark scope must also be read separately: the MLE-bench Lite lead is relative to all comparison models listed in the report, the Frontier-CS lead applies only within open-weight models that report a score for that item, and the BrowseComp and HLE "highest" claims are limited to models with no more than 40B parameters — the three statements cannot be merged into a cross-scale overall victory. The capability direction points to long-turn tasks with verifiable feedback (machine learning experiments, algorithmic problems), which is more likely to transfer to engineering scenarios with test or metric feedback, such as code repair and data pipeline tuning; in open-ended writing or vague requirements without objective feedback, "reflection" may just be multiple rounds of spinning while tokens and time are still consumed. In addition, all scores are self-reported by the team, and the report shows no independent third-party reproduction; the comparison set includes both closed frontier models and models whose parameter scale differs by orders of magnitude. The existence of a weights link indicates high accessibility, but having weights is not the same as an open-source license, and the model card and license text must be checked before commercial use.
Integration and evaluation
The following is advice and conditional analysis, not reported fact. First, if your workload has automatically judgeable feedback signals (unit tests, metric thresholds, a hidden evaluation set), start with small-scale self-deployment validation: choose a quantized version that fits existing VRAM, wire each round of feedback into the agent loop, and use "the slope of pass rate against turns under a fixed turn budget" as the metric; if the slope flattens within 3 to 5 turns, long-horizon reflection offers limited benefit for your task, and a more direct approach is better. The premise is that you have confirmed the license permits your use and can bear the operational cost of a self-built inference stack. Second, if you care more about scores on retrieval research tasks like BrowseComp, the decision point should be the "turn limit for multi-round retrieval and reflection" rather than a single-point score: on an internal retrieval QA set, fix the same retriever and tool set, compare the success rate per unit task cost for "turn limit 2" versus "turn limit 8" configurations, and if the success-rate gain is smaller than the increase in inference cost, fall back to the low turn count. Third, if your scenario is open-ended generation without objective feedback or strongly depends on a stable low-latency online interface, do not replace your current model on the basis of the report alone: first compare against your current model on the same batch of internal samples, record the tokens and time consumed per reflection round, and judge by success rate per unit task cost; whichever path you take, read the evaluation setup, turn budgets and scoring mechanisms in the paper rather than just the single-point numbers.
Limits and open questions
The report does not disclose AREX-2's license or commercial terms, nor its context length, inference overhead or any hosted API; all four scores are self-reported by the team, with no third-party reproduction seen, and the comparison scopes differ (all comparison models listed in the report, open-weight models reporting a score for that item, and models with no more than 40B parameters), so they cannot be merged into one conclusion across scopes. Training data and evaluation details need to be verified against the paper. It should also be clear that AREX-2 currently has no exact match in the KimiSeek model catalog, which lists only its base, Qwen3.8 27B; the two are different versions and cannot be equated. The report mentions no API availability for this model, and any such information should follow the publisher's statements.