The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestBAAI2
BAAIReasoning

AREX-2 arrives as an open deep-research agent model

The reasoning-focused release pairs tool use with long context and self-improvement, and ships as a vision-language model on Hugging Face.

Sep 29, 2026
NotableOther
AREX-2

A new model called AREX-2 has appeared on Hugging Face, positioned as a "deep-research" agent built around reasoning, tool use, long context and what its authors describe as self-improvement. The release is hosted under the BAAI organization at huggingface.co/BAAI/AREX-2 and is tagged for both reasoning and vision-language workloads.

The framing matters because "deep research" has become shorthand for a specific class of system: one that can plan a multi-step investigation, call external tools, and read long documents before producing an answer. Rather than a single-shot chat model, AREX-2 is pitched at that agentic workflow, where the model decides when to retrieve, compute, or verify.

What the record tells us

  • A reasoning-first model with vision-language (VLM) capabilities
  • Designed for tool use and long-context tasks
  • Described as supporting self-improvement in its research loop
  • Published under an "other" license, so terms should be checked before use

Several specifics remain unstated at launch. The listing does not disclose a parameter count, context-window length, or benchmark results, and it is marked as a dense (non-MoE) model. Those details will determine how AREX-2 compares with the growing field of open agentic systems.

For now, the release is most useful as a signal of where open models are heading: away from standalone answer engines and toward configurable agents that combine reasoning, perception and tools. Teams interested in evaluating it should review the model card directly, particularly the license and any usage requirements, before building on top of it.

Sources

  • BAAI/AREX-2

    Hugging Face

    Visit
OlderXiaomi Enters Document OCR With New Vision ModelXiaomi · Vision-Language · 2 months agoNewerJEV-27B-VL Pairs Vision-Language With Calibrated OddsAutotrust · Vision-Language · 2 months ago

Get the model

Hugging Face

Specs

Context window262K tokens
Size54.7 GB
PrecisionBF16
ArchitectureQwen3_5ForConditionalGeneration
LicenseOTHER
Downloads3.9K
Likes219

Can you run it?

Runs on a 24 GB GPU at 4-bit.

  • BF16 (as published)

    80 GB GPU (A100 / H100) · 128 GB Mac Studio

    57.8 GB
  • 8-bit

    48 GB GPU (RTX 6000) or 2×24 GB · 64 GB Mac

    32.2 GB
  • 4-bit

    24 GB GPU (RTX 3090 / 4090) · 32 GB Mac

    19.6 GB
Your machine
GGUF builds
  • KV cache sized for 8,192 tokens of context; longer prompts need proportionally more.

Based on Hugging Face weights metadata. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

ReasoningVision-Language

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More from BAAI

All BAAI releases →
Emu3.5
BAAI/Any-to-AnyConsumer GPU

BAAI Releases Emu3.5, an 'Any-to-Any' Multimodal Model

The new open-source model from the Allen Institute for AI unifies text and image understanding and generation into a single architecture.

Oct 31, 2025

More in Reasoning

All Reasoning →
Sage S2.1 Flash
Internlm/ReasoningMulti-GPU server

InternLM's Sage S2.1 Flash targets scientific reasoning

The 320B mixture-of-experts model is built to reason over scientific evidence and act inside interactive environments.

Nov 15, 2026
ServiceNow/Text → SpeechConsumer GPU

ServiceNow Debuts Apriel Voice 27B, an Open Speech Model

The 27-billion-parameter preview unifies text-to-speech, transcription, and reasoning in a single open model.

Oct 28, 2026
SenseNova-Loop-9B
SenseTime/Vision-LanguageRuns on a laptop

SenseTime previews a 9B looped-transformer VLM

SenseNova-Loop-9B-Preview puts a recurrent-depth architecture to work on multimodal reasoning.

Oct 28, 2026