A 744B-parameter research agent that closes the loop from question to result

What it does

Built by the Shanghai Artificial Intelligence Laboratory on a 744B-parameter MoE GLM-5.2 foundation, Atria Dawn Preview is a text-only agentic model with a 256K context window. It targets scenarios requiring continuous environmental understanding, tool use, and multi-step task completion across discovery, creation, delivery, and cybersecurity. The model attempts to close the full loop from problem analysis and solution design through code implementation, experiment execution, and failure recovery.

The interesting bit

While most LLMs generate answers, this preview is explicitly optimized for “end-to-end delivery” in real-world productivity scenarios like scientific automation and office work. The README also includes unusually detailed workarounds for forcing text-only mode in Codex and Claude Code, which is a practical admission that the model rejects any multimodal input outright.

Key highlights

• 744B-parameter MoE architecture with 256K context window, available in full and FP8-quantized weights
• Competitive on search and agentic benchmarks: 96.0 on DeepSearchQA, 92.5 on BrowseComp, and 86.5 on CyberGym
• Supports both local inference via SGLang and vLLM and hosted API access with regional endpoints
• Explicitly text-only: no image, PDF, or binary input support
• Weights available through both Hugging Face and ModelScope

Caveats

• The README contains commented-out local deployment instructions and a truncated Claude Code integration example
• It does not lead every benchmark; Claude Opus 5 and GPT 5.6 sol outperform it on several coding and delivery tasks
• Strictly text-only input limits utility for workflows requiring document or image analysis

Verdict

Worth evaluating if you need a high-context agent that can iterate with tools and execute experiments, particularly in scientific research or security workflows. Skip it if you require multimodal input or a model that dominates every coding benchmark.