BAAI-AREX-Turbo-GGUF

AREX-Turbo (BAAI/AREX-Turbo) is an Apache-2.0 open-weight 4B dense deep-research agent from BAAI, built on Qwen/Qwen3.5-4B as a compact alternative to the 122B AREX-Base, with a 256K context window. Its recursive self-improvement framework combines iterative research, evidence verification, targeted follow-up searches, and autonomous context updates to improve long-horizon information seeking. Across research benchmarks, it scores 70.7 on BrowseComp, 81.6 on GAIA, 57.0 on xbench-2510, 78.5 on DeepSearchQA, 68.5 on WideSearch-en, and 40.6 on text-only HLE with tools. Despite its smaller size, it outperforms several 30B–235B models, matches Qwen3.5-122B on GAIA, surpasses it on BrowseComp (70.7 vs. 63.8) and WideSearch-en (68.5 vs. 60.5), and exceeds Gemini-3.1-Pro on GAIA, xbench-2510, and WideSearch-en. It uses search, visit, update_context, and finish tools for evidence-driven research.

[!NOTE] MTP — Speculative Decoding: No

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp