Shanghai AI Lab releases Atria Dawn, a 512K-context agentic model built on 744B-parameter GLM-5.2
Shanghai Artificial Intelligence Laboratory has released Atria Dawn, a text-only agentic model with a 512K context window, built on the 744B-parameter MoE GLM-5.2 foundation model. The lab reports results on 25 benchmarks covering search, coding, tool use, productivity and security. Pricing is not yet disclosed.
Shanghai Artificial Intelligence Laboratory has released Atria Dawn, an agentic model with a 512K-token context window, built on the 744B-parameter mixture-of-experts GLM-5.2 foundation model. The model is published under the internlm organization on Hugging Face, with weights also listed on ModelScope.
What Atria Dawn targets
According to the lab, Atria Dawn is aimed at research and engineering scenarios that need continuous environmental understanding, tool use and multi-step task completion. The lab says it supports the full loop of problem analysis, solution design, tool use, code implementation, experiment execution, result analysis and failure recovery. It describes four focus areas:
- Discovery: retrieving and organizing evidence, deep research, and turning research questions into experimental plans.
- Creation: building software, interactive applications, games, data visualizations and machine learning systems.
- Delivery: turning documents, data and design requirements into reports, presentations and other structured deliverables.
- Cybersecurity: analyzing issues, validating vulnerabilities, applying fixes and re-validating in authorized environments.
Releases and specs
Four checkpoints are listed, all Instruct models:
| Model | Context |
|---|---|
| Atria Dawn (full release) | 512K |
| Atria Dawn FP8 (full release) | 512K |
| Atria Dawn Preview | 256K |
| Atria Dawn Preview FP8 | 256K |
Atria Dawn accepts text input only. Per the model page, the endpoint rejects image input with a 400 error. Pricing is not yet disclosed.
Benchmarks (lab-reported)
The lab compared Atria Dawn and Atria Dawn Preview against DeepSeek V4 Pro 0813, KIMI K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol and Claude Opus 5. These figures come from the lab's own table, which it says matches its official website.
Selected Atria Dawn scores (Preview in parentheses):
- AutomationBench: 54.7 (53.8)
- CyberGym: 86.8 (86.5)
- SkillsBench: 67.8 (66.4)
- WideSearch: 84.6 (81.9)
- MLE-bench Lite: 88.2 (86.2)
- PaperBench: 89.6 (87.0)
- BrowseComp: 91.3 (92.5)
- DeepSearchQA: 96.5 (96.0)
- BFCL V4: 78.4 (77.0)
- Terminal-Bench 2.1: 81.0 (78.3)
- SWE-bench Pro: 61.7 (59.6)
On several coding benchmarks the table shows other models scoring higher. Claude Opus 5 is listed at 90.2 on Terminal-Bench 2.1, 76.5 on SWE-bench Pro, 79.9 on NL2Repo-Bench and 41.0 on Harbor-Index, against Atria Dawn's 81.0, 61.7, 50.0 and 17.3. Atria Dawn is listed at 49.3 on DeepSWE, with no Preview result. Other cells in the table are marked unavailable.
Deployment and access
Local deployment requires SGLang v0.5.13.post1 or later, or vLLM v0.23.0 or later. Hosted access is through separate international and China API consoles. Requests use Atria-Dawn as the model value.
The OpenAI-compatible base URL is https://discovery-api.intern-ai.org.cn/v1. The same host without /v1 serves Anthropic-compatible clients and SystemOne. The documentation includes a Codex configuration that uses the Responses API and requires Codex CLI 0.154.0 or later. It also covers a Claude Code setup with a PreToolUse hook that blocks the Read tool on PDF and image files. For Codex, the guide sets input_modalities to text and context_window to 512000 so the client budgets the prompt correctly and does not send images.
What this means
Atria Dawn pairs a 744B-parameter MoE base with a 512K window and ships FP8 variants for both the full and Preview checkpoints. The documented support for SGLang, vLLM, OpenAI-compatible and Anthropic-compatible endpoints, Codex and Claude Code means existing agent tooling can be pointed at it with configuration changes. The text-only limit matters for workflows that depend on screenshots or PDFs, which the documentation tells users to handle by pasting text. On the lab's own numbers, the model's strongest showing is in search, research and security tasks, while its coding results are mixed against the listed comparison models. Independent verification of these scores and pricing details are still outstanding.
Related Articles
Shanghai AI Lab releases Sage-A1.5, a 35B MoE agentic model with 256K native context
Shanghai AI Laboratory has released Sage-A1.5, a 35B-parameter Mixture-of-Experts agentic model with a native 256K context window, extendable to 512K via YaRN. The lab claims it matches or beats larger frontier models on several agentic and scientific benchmarks, including 95.1 on GAIA. Pricing is not yet disclosed.
MiniMax releases M3.1-Flash: 428B-parameter multimodal MoE with 1M-token context and weights on Hugging Face
MiniMax has published MiniMax-M3.1-Flash on Hugging Face: a native multimodal mixture-of-experts model with about 428B total parameters, 23B activated per token, and a context window of up to 1M tokens. It builds on MiniMax-M3 with sparse attention, quantization-aware training and DSpark speculative decoding.
Tencent releases WeVisOne-VL-2B, a 2B-parameter model for object detection, segmentation and alpha matting
Tencent's WeChat Vision team released WeVisOne-VL-2B, a 2B-parameter vision-language model built on Qwen3-VL. It handles object-level and pixel-level perception, including referring segmentation and alpha matting, through structured next-token prediction and a 9.5M-parameter pixel decoder.
inclusionAI releases Ling-3.0-flash-Sante, a 124B healthcare model with 5.1B active parameters
inclusionAI has released Ling-3.0-flash-Sante, a healthcare-specialized model built on Ling-3.0-flash with 124B total and 5.1B active parameters per token. Weights are on Hugging Face and ModelScope, and training extends to 256K tokens. The company labels it a research model, not a medical device.
Comments
Loading...