BAAI Put AREX-2 on Hugging Face Last Night, and the Repo Is Nearly Empty

At 17:56 UTC on September 29, 2026, the Beijing Academy of Artificial Intelligence created a Hugging Face repository called AREX-2. As of this writing that repository contains one file — a .gitattributes— and it stores zero bytes of model data. There is no model card, no parameter count, no architecture description, no benchmark, no licence, no announcement on BAAI's own site, and no social post from anyone at the lab. Hugging Face's API reports zero downloads and zero likes. No inference provider serves it. BAAI has said nothing at all about it. This is not a launch story. It is the story of a name being reserved by a lab that does not reserve names casually, published here because the one thing that is knowable — who put the repo up, and what that org's 2026 record looks like — is enough to say something useful about what is probably coming, and to be precise about the difference between that and knowing.

The distinction matters more than usual here. A repo created a few hours before this sentence was written is genuinely recent, and a what-we-know-so-far piece about it can be accurate. What it cannot do is describe a model. So everything below is split cleanly: facts read directly off Hugging Face this session, and expectations derived from the AREX family that already shipped. If you have come here for AREX-2's specs, the honest answer is that AREX-2 does not have any yet, and a paragraph of plausible-sounding numbers would be worse than this one.

What is actually in the repository

Everything publicly knowable about AREX-2 fits in a short list, and it is worth stating in full so that nothing further down gets mistaken for it.

• The repo exists — huggingface.co/BAAI/AREX-2, created 2026-09-29T17:56:21Z, last modified at the same instant. A single commit, at creation, is the whole git history.

• It is not empty of intent — the sibling name it points at is not a random string. BAAI's AREX line already exists on that same organisation: AREX-Base and AREX-Turbo, both created on 23 July 2026, both full weights drops under Apache 2.0.

• It is empty of everything else — no weights, no config, no tokenizer, no README prose, no licence tag, no downloads, no likes, no provider deployment.

• It is brand new, not resumed — the repo has a single creation timestamp. There is no earlier incarnation of this name under BAAI, and no re-upload under a variant spelling that turned up in a search.

The screenshot below is the entire public surface of it. That is the fact the rest of this piece has to stay faithful to: BAAI has claimed a name, and has not yet put anything behind it.

What "AREX-2" most likely continues

The reason a near-empty repo is worth eighteen hundred words is that the thing it is probably a second version of is unusually well documented.

AREX-Base, from July, is a 122-billion-parameter mixture-of-experts model with 10 billion active parameters, built on a Qwen3.5-122B-A10B base and released under Apache 2.0. It is not a chat model that happens to search well; it is a deep-research agent with a two-loop design. An inner loop gathers evidence, assembles a candidate answer and attaches a confidence figure to it. An outer loop checks that answer against the original constraints and can accept it, refine it, or throw the trajectory away and start again. Between the loops it maintains a context block holding verified findings, open candidates, unresolved constraints and the next plan, and it drives search and browsing as tools.

On BAAI's own published figures — vendor-reported, not independently reproduced — that agent scores 82.5 on BrowseComp, 85.4 on GAIA, 71.0 on xbench-2510, 89.9 on DeepSearch QA and 82.0 on WideSearch-en. Its 4B sibling AREX-Turbo, built on Qwen3.5-4B, scores 70.7, 81.6, 57.0, 78.5 and 68.5 on the same set. Those are strong numbers for a first-generation agent, and the gap between the 122B and the 4B is the interesting part of the family: BAAI built a quality tier and a serving-cost tier in the same release.

None of that transfers automatically to AREX-2. A "2" in a model name is a naming convention, not a contract. It could mean a second-generation architecture, a larger Base, a smaller distillation, a retrained agent with the two-loop framework kept and the backbone swapped, or something the family name does not predict at all. What it does suggest — and this is inference, plainly labelled — is that BAAI is continuing an agent line rather than opening a new one, because a lab that wanted a fresh model would not spend its own family name on it.

Where the rest of BAAI's 2026 record sits

The strongest thing that can be said about AREX-2 today is about the org, not the model. BAAI's 2026 Hugging Face history is a steady line of real weights drops: Orca-4B on 12 July, the AREX-Base and AREX-Turbo pair on 23 July, then Recon2Reason-Reasoning-4B, ConsiSpace, AIDD and Brainmu-Spike through September, plus a longer tail across the RoboBrain, URSA, Emu3.5 and OpenSeek lines. They are not all famous, and they are not all benchmarked, but they are all shipped artifacts under open licences.

That pattern is why this repo is worth a post and a squatted name from an anonymous account would not be. It is also the reason to be careful: the same org has, in the last four weeks, put up repositories that stayed empty longer than a day. A reserved name is a signal about intent, and intent is not a release. If you are planning capacity around AREX-2, you do not have a date and you should not invent one.

What would have to happen for this to become callable

There are four checkpoints between "repo exists" and "you can put it in production", and they can arrive in any order.

The first is weights: actual safetensors in the tree, which is the point at which the model becomes inspectable and quantisable by anyone. The second is a card: the parameter count, base model, context window and evaluation table that tell you what you are downloading. The third is a licence, which for this family has been Apache 2.0 — a permissive default that, if it holds, removes the revenue-capped licensing question entirely.

The fourth is a route, and this is the one most readers of this blog actually care about. AREX-Base is not on OrcaRouter's catalogue today, and neither, obviously, is AREX-2. When a model like this does land with open weights, the useful thing about a routing layer is that getting to it does not require a second contract or a new client — one API in front of 200-plus models, provider list price passed through with no markup added, so a vendor's own price at the moment of publication is the price you see, and automatic failover so a single provider's bad afternoon does not become your agent's bad afternoon. For a research agent that fans out across dozens of tool calls per query, that last property is worth more than it sounds: the failure modes of a deep-research loop are mostly downstream of one flaky provider.

We will say the same thing when AREX-2 has weights that no host is serving yet: failover is how you test an unproven model on a side path without betting the main one on it.

What this page will look like when the repo fills in

Three outcomes are possible, and they are distinguishable from outside.

If the tree gains safetensors in the next few days, this becomes a genuine quiet-ship story and gets rewritten with real parameter counts and the licence read off the card. That is the common shape for this organisation and it is the outcome the timestamp points at.

If the repo sits empty for weeks and a paper appears first — the AREX family's first generation shipped alongside arXiv 2607.21461 — then the honest headline is that a second generation is in research and the repo was a placeholder for it.

If nothing happens at all, then nothing happened, and the correct thing to publish is nothing. A reserved name that never fills in is not a story, and it should not be dressed up as one.

Two facts are enough to check this yourself, without relying on anyone's summary. The repository's file tree, which either has weights in it or does not, and the repository's creation date, which will not change. Everything else about AREX-2 — its size, its benchmarks, whether it is even called AREX-2 when it ships — is currently unknown, and this piece will not pretend otherwise.