跳转到主要内容

Atria Dawn Preview

Shanghai AI Laboratory开源权重MIT · 商用许可

描述

Atria Dawn Preview is a preview-stage agentic model from Shanghai AI Laboratory (published under the internlm Hugging Face org), built on a 744B-parameter MoE foundation (the model card states it is trained on the GLM-5.2 base). It targets research and engineering workflows requiring sustained environment understanding, tool use, and multi-step task completion across four capability areas the vendor calls Discovery, Creation, Delivery, and Cybersecurity. The Instruct model accepts text-only input (image/PDF input returns an API error) with a 256K-token context window. Weights are released under the MIT license on Hugging Face and ModelScope in BF16 and FP8 formats. It is served via an international API at api.atria-asi.ai (Chat Completions, Messages, and Responses wire formats) and a China-region endpoint; no per-token pricing has been published on either endpoint as of this writing.

发布日期
2026-09-11
参数规模
744.0B
上下文长度
支持模态

能力雷达图

100
general
70
coding
100
reasoning
85
science估算
60
agents
0
multimodal

缺少专门科学评测时,Science 由 LLM Stats 科学得分或推理能力估算。

排行榜排名

领域#排名分数来源
智能体能力模型榜2
75.0
LS

基准测试分数 (LLM Stats)

(LLM Stats (zeroeval))

Agents

AutomationBench53.8%自报
Tau3 Banking41.2%自报

Code

CyberGym86.5%自报
MLE-Bench Lite86.2%自报
SkillsBench66.4%自报

General

GDPval-AA1583.00 / 3000自报

Reasoning

BrowseCompOpenAI (2025)92.5%自报
Terminal-Bench 2.178.3%自报
SWE-Bench ProPrinceton NLP (2024)59.6%自报

Search

DeepSearchQA96.0%自报
WideSearch81.9%自报

Tool Calling

BFCL-V477.0%自报

AA 评测指数

(Artificial Analysis)

暂无 AA 评测数据

LLM Stats 分类评分

(LLM Stats (zeroeval))
Legal
100
Finance
100
Reasoning
100
General
100
Agents
100
Safety
90
Search
90
Code
70
Tool Calling
60

定价

暂无定价数据

速度

暂无速度数据

供应商价格排行

暂无提供商数据

外部链接