Shanghai AI Lab Releases Atria Dawn Preview, a 744B-Parameter MoE Agentic Model Built on GLM-5.2
Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, a text-only agentic model built on the 744B-parameter MoE GLM-5.2 foundation model with a 256K context window. The model targets multi-step research, coding, and productivity tasks, with benchmark results claimed to compete with DeepSeek V4 Pro, Kimi K3, and Claude Opus 5.
Atria Dawn Preview — Quick Specs
Shanghai AI Lab Releases Atria Dawn Preview, a 744B-Parameter Agentic Model
Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, a new agentic model built on the 744-billion-parameter mixture-of-experts (MoE) GLM-5.2 foundation model. The model is available now on Hugging Face and ModelScope, published under the internlm organization, and supports a 256,000-token context window.
Atria Dawn Preview is a text-only instruct model — it does not accept image, PDF, or other binary inputs. The lab ships two versions: a standard instruct model and an FP8-quantized variant, both with the same 256K context limit. Pricing for hosted API access has not been disclosed.
Designed for multi-step agentic work
According to Shanghai AI Lab, the model is built for "continuous environmental understanding, tool use, and multi-step task completion," combining task objectives with environmental feedback across a full loop of problem analysis, tool use, code implementation, experiment execution, and failure recovery. The lab frames the model's capabilities across four use-case categories: Discovery (research and evidence retrieval), Creation (software and ML system building), Delivery (turning data into reports and presentations), and Cybersecurity (vulnerability analysis and remediation in authorized environments).
Benchmark claims
Shanghai AI Lab published comparisons against DeepSeek V4 Pro 0813, Kimi K3, Qwen3.8 Max, GLM 5.3, GPT-5.6, and Claude Opus 5 across search, coding, tool-use, productivity, and security benchmarks. According to the company's own reported figures:
- BrowseComp: 92.5 (vs. 83.4 for DeepSeek V4 Pro, 91.2 for Kimi K3)
- AutomationBench: 53.8 (highest among all listed models)
- BFCL v4 (tool use): 77.0 (highest among all listed models)
- CyberGym: 86.5 (highest among all listed models)
- SWE-bench Pro: 59.6 (below Kimi K3's 61.6 and Claude Opus 5's 74.7)
- Terminal-Bench 2.1: 78.3 (trailing Qwen3.8 Max at 89.3 and Claude Opus 5 at 90.2)
These are the lab's self-reported numbers; none have been independently verified. Atria Dawn Preview leads on several tool-use and search benchmarks but trails competitors, notably Claude Opus 5, on coding-heavy tasks like SWE-bench Pro and Terminal-Bench.
Deployment
The model can be run locally via SGLang (v0.5.13.post1+) or vLLM (v0.23.0+), or accessed through hosted API endpoints in international and China regions. Shanghai AI Lab has also published integration guides for Codex, Claude Code, and Kimi Code, including configuration steps to disable multimodal input handling in client tools that assume image support by default — a necessary step since the model rejects image inputs with a 400 error.
What this means
Atria Dawn Preview is a preview-stage release aimed squarely at agentic workflows — research, coding, office productivity, and security testing — rather than general chat or multimodal use. Built on a 744B-parameter MoE base, it competes directly with frontier-scale models from DeepSeek, Moonshot (Kimi), Alibaba (Qwen), Zhipu (GLM), OpenAI, and Anthropic on agentic benchmarks. The self-reported scores show strength in tool-calling and browsing tasks but weaker performance on demanding coding benchmarks compared to Claude Opus 5, suggesting the model may be better suited to research and automation pipelines than production software engineering. As a preview release with text-only input, it should be treated as an early checkpoint rather than a finished product — independent verification of the benchmark claims is not yet available.
Related Articles
China Telecom Releases Xing4.0-29B-A4B, a 29B MoE Model Trained Entirely on Ascend NPUs
China Telecom Artificial Intelligence Technology has released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts model with only 4B parameters active per token and native 256K context. The company claims it is the first model of this scale trained entirely on Huawei's Ascend NPU platform using the MindSpore framework.
Unbiased Launches Pareto, a $2.50/$7.50-per-Million-Token Multimodal Model for Coding and Agents
Unbiased has released Pareto, a multimodal composite model aimed at research, coding, and agentic workflows. The model offers a 262K context window and is priced at $2.50 per million input tokens and $7.50 per million output tokens via OpenRouter.
PrismML's Bonsai 2 Compresses 27B-Parameter Model to 5.9GB, Retains 98% of Benchmark Performance
PrismML released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B model that shrinks memory footprint by 9x to 10x down to 5.9GB. The startup claims 98% aggregate benchmark parity with the original, up from 95% in its first release, using a ternary weight compression technique.
OpenAI's GPT-6 Astra Beats Pokémon in 18 Hours, Scores 62.7% on ARC-AGI-3
GPT-6 Astra completed Pokémon FireRed in 18 hours 12 minutes, five times faster than its predecessor, and scored 62.7% on ARC-AGI-3 versus 7.78% for GPT-5.6 Sol. The model also ran a 141-hour Minecraft session and finished Fallout 3 in roughly 59 hours, according to independent testers.
Comments
Loading...