Agent 与自动化 5.0 · 必读 2026-09-11 · 产品

Atria Dawn Preview Open-Source 744B MoE Agent Weights (internlm/Atria-Dawn-Preview)

Atria Dawn Preview 是上海 AI Lab / InternLM 线在 Hugging Face internlm/Atria-Dawn-Preview 开源的 744B MoE 长程 agent 预览版模型卡,2026-09-11 创建权重仓09-12 出 FP8 变体...

打开原文回到归档

Atria Dawn Preview — Open-Source 744B MoE Agent Weights (internlm/Atria-Dawn-Preview)

HF Card 摘要

Hugging Face internlm/Atria-Dawn-Preview 模型卡(HF API 200,likes ≈54,created 2026-09-11):744B 参数 MoE,基座 GLM-5.2,上下文 256K,许可 MIT。FP8 变体 2026-09-12 上线;ModelScope 同步权重。发布定位:From Research Questions to Verifiable Results。作者 harness(未独立复测):AutomationBench 53.8 / BrowseComp 92.5 / CyberGym 86.5 / DeepSearchQA 96.0 / Workspace-Bench-Lite 68.2 / SkillsBench 66.4 / BFCL v4 77.0。演示叙事层:天气预测网络(无外网检索)、约 20 分钟搭 MiniOS、授权环境下找洞-利用-修补-再验证流程——与卡表数字属不同叙述层。

仓库要点 (中文)

上海 AI Lab / InternLM 线在 Hugging Face internlm/Atria-Dawn-Preview 给出 Atria Dawn Preview 权重仓,2026-09-11 创建、09-12 增 FP8 变体;MIT 许可。模型卡把发布定位写成「From Research Questions to Verifiable Results」,强调长程任务、工具调用、失败恢复、可复现结果;评测表是作者 harness 的官方数字,不等于第三方复测。HF likes 与 GitHub stars 持续上行(窗口约 85 / 242,CST 21:00)。本地/代理推理先核 FP8 与体量(BF16 约 1.5TB 量级),写稿把演示故事和卡表分数分开。

Key claims (English)

Atria Dawn Preview is the open-weight release from Shanghai AI Lab / InternLM on Hugging Face at internlm/Atria-Dawn-Preview (HF API 200, likes ≈54, created 2026-09-11), with a 744B-parameter MoE, 256K context, and MIT license. FP8 variant went live 2026-09-12; ModelScope mirrors the weights. The card positions the release as 'From Research Questions to Verifiable Results', emphasizing long-horizon tasks, tool use, failure recovery, and reproducible results. The scoreboard (AutomationBench 53.8, BrowseComp 92.5, CyberGym 86.5, DeepSearchQA 96.0, Workspace-Bench-Lite 68.2, SkillsBench 66.4, BFCL v4 77.0) is the author's harness, not a third-party replication. Demo narratives (a weather-prediction net built without web retrieval, a 20-minute MiniOS bootstrap, an authorized exploit–patch–re-verify loop) are explicitly a separate layer than the scoreboard. For local / on-prem inference, audit FP8 and weight size first (BF16 class ≈ 1.5TB) and keep demos separate from scoreboard claims.

Obsidian 证据摘录

「团队:上海人工智能实验室;HF 组织 internlm;仓 atria-asi/Atria-Dawn-Preview(API stars≈89)。基座:744B 参数 MoE GLM-5.2;上下文 256K;许可 MIT。权重:internlm/Atria-Dawn-Preview(HF likes≈54,created 2026-09-11)+ FP8 变体(09-12)。」——Hermes 定时任务/X-每日简报/2026-09-14-X-Hot-Brief_AtriaDawn-OM1-GlassImaging.md L35-37

链接