Atria Dawn Preview
Released Sep 11, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Modality
- Text
- Context window
- 262K
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Benchmarks2 of 10 reported
SWE-bench Pro
Terminal-Bench 2.1
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Built for the research loop rather than chat — it searches literature, calls tools, runs the experiment, reads the result and iterates, and its strongest results are all on discovery and tool-use suites
- Recovers from a failed tool call by re-planning instead of stalling, which is where the post-training went: a verifiable-experience pipeline that ties every tool interaction to an executable environment and a checked outcome
- MIT weights with SGLang and vLLM support from day one, so self-hosting is a licensing non-event for anyone with the hardware
- Sparse attention inherited from the GLM-5.2 base cuts the KV cache roughly fourfold, which is what keeps the 256K window serviceable on a 744B model
Weaknesses
- Text only: images, PDFs and audio are rejected with a 400 at inference time, so anything visual has to be pre-processed by another model
- Tuned for discovery and tool use, not software engineering — coding is where it lands furthest from the closed flagships, and it is not the model to reach for on a pure SWE task
- Every benchmark is vendor-reported with no independent evaluation yet, and the 1.5 TB BF16 (756 GB FP8) footprint puts local verification out of reach for most teams
Shanghai AI Laboratory's first dedicated agentic foundation model, released under the Atria brand (atria-asi.ai; Hugging Face org internlm): agentic post-training on Zhipu's 744B-parameter GLM-5.2 MoE base through a Verifiable Experience Pipeline. Weights went up on Hugging Face on 2026-09-11 with no announcement, the FP8 checkpoint on 09-12, the 140-author technical report (arXiv 2609.15818) on 09-14 and a press release on 09-15. SWE-bench Pro 59.6 and Terminal-Bench 2.1 78.3 are from the model card's own 16-benchmark table, which also reports DeepSearchQA 96.0, BrowseComp 92.5, WideSearch 81.9, DeepResearch Bench II 51.1, MLE-bench Lite 86.2, BFCL v4 77.0, AutomationBench 53.8, SkillsBench 66.4, tau3-Bench Banking 41.2, Workspace-Bench 65.0, JobBench 50.3, CyberGym 86.5 and a 'GDPval' Elo of 1583 whose version is unstated and whose comparison-model figures match neither Artificial Analysis' v2 nor v2.1 boards, so gdpvalAA is null. GPQA Diamond, HLE, SWE-bench Verified, ARC-AGI-2 and LMArena are unpublished, and AA has no page for it. The active-parameter count is not on the card; third-party guides put it at 18B over 256 experts and 78 layers, inherited from GLM-5.2. The international API (api.atria-asi.ai, model id Atria-Dawn-Preview, Google sign-in) has published no price and LLM Gateway serves it free at launch, hence availability general with pricing null. Context is the card's 256K (262,144 per hosts); max output 65,536. predecessorId is null: this is the lab's first Atria model and it is not framed as replacing any InternLM release.
For developers
API model strings
- atria
Atria-Dawn-Preview
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor