{"schema":"postcutoff/event@1","as_of":"2026-10-17T22:10:00+02:00","url":"https://postcutoff.com/e/2026-09-11-shanghai-ai-lab-atria-dawn-preview/","md":"https://postcutoff.com/e/2026-09-11-shanghai-ai-lab-atria-dawn-preview/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-09-11-shanghai-ai-lab-atria-dawn-preview","date":"2026-09-11","date_precision":"day","short_title":"Atria Dawn Preview: Shanghai AI Lab's 744B open agent model","deck":null,"takeaway":"Its own table puts it first on BrowseComp (92.5), DeepSearchQA, BFCL v4, AutomationBench and CyberGym, but behind Claude Opus 5 on SWE-bench Pro and JobBench.","category":"open-source","category_label":"Open source","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Hugging Face: internlm/Atria-Dawn-Preview (model card)","url":"https://huggingface.co/internlm/Atria-Dawn-Preview","type":"official","group":"primary","domain":"huggingface.co"},{"n":2,"title":"Atria API docs","url":"https://api.atria-asi.ai/docs","type":"docs","group":"primary","domain":"api.atria-asi.ai"},{"n":3,"title":"Pandaily: Shanghai AI Laboratory releases Atria Dawn full version with a 512K context window (Oct 17)","url":"https://pandaily.com/shanghai-ai-lab-atria-dawn-full-release-512k-context-jobbench-weights-coming-soon","type":"press","group":"press","domain":"pandaily.com"},{"n":4,"title":"Startup Fortune: Shanghai AI Lab quietly releases Atria, a free 744 billion parameter AI agent","url":"https://startupfortune.com/shanghai-ai-lab-quietly-releases-atria-a-free-744-billion-parameter-ai-agent/","type":"press","group":"press","domain":"startupfortune.com"}],"official":2,"filed":"2026-10-17","updated":"2026-10-17","orgs":["Shanghai AI Laboratory"],"title":"Shanghai AI Laboratory releases Atria Dawn Preview, a 744B MIT-licensed open-weight agent model built on GLM-5.2","summary":"On Sept 11, 2026 Shanghai AI Laboratory posted Atria Dawn Preview on Hugging Face: an MIT-licensed agentic model built on Z.ai's 744B-parameter MoE GLM-5.2, with a 256K context, aimed at long research, coding, office and security tasks. Its own table puts it first on BrowseComp (92.5), DeepSearchQA, BFCL v4, AutomationBench and CyberGym, but behind Claude Opus 5 on SWE-bench Pro and JobBench. On Oct 17 Pandaily reported a full version with a 512K context and weights \"coming soon\".","key_facts":["Hugging Face repo internlm/Atria-Dawn-Preview created Sept 11, 2026; MIT license; ~753B parameters in safetensors; text-only input; 256K context; FP8 build Sept 12 and an Ascend w8a8 build Sept 15; also on ModelScope","Base: 744B-parameter MoE GLM-5.2 (Z.ai); model card cites arXiv 2609.15818","Vendor benchmarks (model card): BrowseComp 92.5 (GPT-5.6 Sol 92.2, Claude Opus 5 90.8); DeepSearchQA 96.0; BFCL v4 77.0; AutomationBench 53.8; CyberGym 86.5; SWE-bench Pro 59.6 (Claude Opus 5 74.7); Terminal-Bench 2.1 78.3 (Opus 5 90.2); GDPval 1583 (Opus 5 1768); JobBench 50.3 (Opus 5 68.0)","Access: API at api.atria-asi.ai (international) and Intern Discovery token plan (China); local serving with SGLang ≥0.5.13.post1 or vLLM ≥0.23.0; the card documents Codex CLI setup via the Responses API","Full version (Pandaily, Oct 17, 2026; not confirmed on the lab's pages yet): 512K context, higher JobBench, WideSearch and PaperBench scores, weights 'coming soon'"],"key_numbers":[],"tags":["open-weights","agents","moe","china","shanghai-ai-lab","glm","mit-license"],"science":null,"body_md":"## What happened\n\nShanghai Artificial Intelligence Laboratory, the state-backed lab behind the InternLM and Intern-S models, published Atria Dawn\nPreview with little fanfare on Sept 11, 2026. It is not trained from scratch: it is post-trained on Z.ai's open 744B MoE\nGLM-5.2. The lab pitches it as an agent that takes \"research questions to verifiable results\", covering discovery (deep\nresearch), creation (software, ML systems), delivery (reports, slides) and cybersecurity in authorised environments.\n\nOn the lab's own table it leads the search and tool-use benchmarks it reports, roughly ties GPT-5.6 Sol on BrowseComp, and trails\nClaude Opus 5 by wide margins on SWE-bench Pro, Terminal-Bench 2.1, GDPval and JobBench. These are vendor numbers.\n\n## Why it matters\n\nIt shows Chinese labs building frontier agent models on each other's open weights: a government lab fine-tuning a startup's\n744B base and releasing it under MIT, with its own API and documented use inside OpenAI's Codex CLI.","disputed":[],"related":[{"id":"2026-10-15-ghys-conjecture-codimension-one-intern-s2-lean","url":"https://postcutoff.com/e/2026-10-15-ghys-conjecture-codimension-one-intern-s2-lean/","date":"2026-10-15","date_precision":"day","short_title":"Ghys's 1995 conjecture proved in the codimension-one case with Shanghai AI Lab's Intern-S2","deck":"Peking University formalizes it in Lean (~550k lines)","takeaway":"Jiesong Zhang's preprint proves that every codimension-one holomorphic Anosov diffeomorphism of a compact complex manifold is conjugate to a hyperbolic torus automorphism, settling Ghys's conjecture in that case and in complex dimension three. It declares that an internal Intern-S2 helped develop the ideas; a Peking University team formalized the main theorems in Lean with its SHEAF agent.","category":"science","category_label":"Science & math","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Result confirmed"]},"sources":7,"official":4,"filed":"2026-10-15","updated":"2026-10-15","orgs":["Shanghai AI Laboratory","Peking University"]},{"id":"2026-10-15-stepfun-step-5-preview-open-weights","url":"https://postcutoff.com/e/2026-10-15-stepfun-step-5-preview-open-weights/","date":"2026-10-15","date_precision":"day","short_title":"StepFun releases Step 5 Preview open weights","deck":null,"takeaway":"Step 5 Preview is one of the largest openly downloadable agent models from a Chinese lab, a day after Ant's 560B Ling 3.1 Flash (MIT).","category":"open-source","category_label":"Open source","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":6,"official":5,"filed":"2026-10-15","updated":"2026-10-15","orgs":["StepFun"]}],"people":[],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-17","type":"filed","text":"Created (found while checking the Oct 17 Pandaily report; the September preview had not been covered)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-17","run":null,"sources_read":null,"updated":"2026-10-17","human_review":null,"version":null},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":134,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":73,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":164,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":103,"in_training_data":false}],"short_url":null}