InternLM Atria Dawn
One-line: The full release of Shanghai AI Lab’s Atria Dawn agentic model (the September Atria-Dawn-Preview was the preview) — 744B MoE built on GLM-5.2, MIT-licensed, 512K context, best-in-row BFCL v4 78.4, CyberGym 86.8, AutomationBench 54.7, DeepSearchQA 96.5 and WideSearch 84.6 against DeepSeek V4 Pro 0813 / Kimi K3 / Qwen3.8 Max / GLM-5.3 / GPT-5.6 sol / Opus 5. ⚠ large — BF16 ≈1.5 TB, first-party FP8 sibling ≈745 GB; server-class only. ⚠ lab-reported numbers only; 0 downloads at capture (launch-day).
Repo created 2026-10-23 03:21Z (16.6h before this run) with the full weight set (368 files) and the first-party FP8 sibling (
internlm/Atria-Dawn-FP8, 192 files) created 2026-10-23 01:23Z. Separate model IDs from the already-tracked internlm—Atria-Dawn-Preview (Preview: 256K ctx, released 2026-09-11). Independent coverage exists (explainx.ai, datastudios.org, Instagram/Twitter launch posts) but no independent benchmark re-run has been found — BenchLM has no page for it.
⚠ All benchmark numbers below are lab-reported from the model card’s comparison table (which the card states matches the official table at atria-asi.ai). No Artificial Analysis / BenchLM / LMArena listing found as of 2026-10-24.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published |
| Aider polyglot | — | not published |
| SWE-bench Verified | — | not published (card uses SWE-bench Pro) |
| SWE-bench Pro | 61.7 | lab — vs Kimi K3 62.0, Qwen3.8 Max 65.3, Claude Opus 5 76.5 |
| SWE-bench Multilingual | — | not published |
| Terminal-Bench 2.1 | 81.0 | lab — vs Qwen3.8 Max 89.6, Opus 5 90.2 |
| DeepSWE | 49.3 | lab — trails DeepSeek V4 Pro 0813 54.9, Kimi K3 70.6 |
| NL2Repo-Bench | 50.0 | lab — trails Kimi K3 63.1, Qwen3.8 Max 63.2 |
| Harbor-Index | 17.3 | lab — trails Kimi K3 24.3, Opus 5 41.0 |
| MLE-bench Lite | 88.2 | lab — vs Qwen3.8 Max 81.3, GPT-5.6 sol 88.9 |
| SpreadsheetBench 2 | 33.0 | lab — vs Kimi K3 35.8 |
| SkillsBench | 67.8 | lab — best in row (vs DeepSeek V4 Pro 65.0, Qwen3.8 Max 66.7) |
| PaperBench | 89.6 | lab — vs Qwen3.8 Max 94.1 |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL v4 | 78.4 | lab — best in row (vs GLM-5.3 74.1, DeepSeek V4 Pro 71.4, Kimi K3 69.1) |
| τ-Bench | τ³-bench Banking 43.3 | lab — vs Qwen3.8 Max 55.2, Opus 5 48.7 |
| ToolACE | — | not published |
| GAIA | — | not published |
| AutomationBench | 54.7 | lab — best in row (vs GLM-5.3 52.5, Qwen3.8 Max 49.7) |
| CyberGym | 86.8 | lab — best in row (vs GLM-5.3 84.5, GPT-5.6 sol 83.6) |
| DeepSearchQA | 96.5 | lab — best in row (vs Kimi K3 95.9, GLM-5.3 94.7) |
| WideSearch | 84.6 | lab — best in row (vs Opus 5 84.0, GPT-5.6 sol 83.3) |
| BrowseComp | 91.3 | lab — vs Kimi K3 91.2, GPT-5.6 sol 92.2 |
| Workspace-Bench / -Lite | 66.1 / 70.0 | lab — -Lite trails GLM-5.3 71.7 |
| JobBench | 57.5 | lab — vs Claude Opus 5 68.0 |
| Harvey LAB | 27.5 | lab — best in row (vs Kimi K3 25.4, DeepSeek V4 Pro 6.7) |
| DeepResearch Bench II | 51.2 | lab — vs Opus 5 54.1 |
| OfficeQA Pro | 67.1 | lab — vs Opus 5 67.2 |
| GDPval | 1594.2 | lab — judge score, vs Opus 5 1714.0 |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published |
| GPQA Diamond | — | not published |
| MATH | — | not published |
| AIME 2025 / 2026 | — | not published |
Efficiency
- Active params: not disclosed (744B-total MoE on GLM-5.2; 256 routed experts, 8 active per token; 78 layers; hidden 6144; vocab 154,880)
- Context length: 512K (card) — the Preview shipped 256K;
max_position_embeddingsin config is 1,048,576 - VRAM: ⚠ BF16 ≈1.5 TB across 368 files; first-party FP8 sibling
internlm/Atria-Dawn-FP8(192 files) ≈745 GB. Not single-80GB-GPU hardware. - Runtime: custom architecture (
model_type: glm_moe_dsa,GlmMoeDsaForCausalLM) — SGLang ≥v0.5.13.post1 or vLLM ≥v0.23.0 using the GLM-5.2 recipes; no stock-transformers path. Text input only (not multimodal). Card ships Codex / Anthropic-compatible / SystemOne client configs plus a hosted API (discovery-api.intern-ai.org.cn) and an international token plan.
What makes it notable
Atria Dawn is Shanghai AI Lab’s agent-native frontier release and the strongest open-weight agentic model on the board by breadth: best in row for BFCL v4 (78.4), AutomationBench (54.7), CyberGym (86.8), DeepSearchQA (96.5), WideSearch (84.6), SkillsBench (67.8) and Harvey LAB (27.5) against a comparison set of DeepSeek V4 Pro 0813, Kimi K3, Qwen3.8 Max, GLM-5.3, GPT-5.6 sol and Claude Opus 5 — while trailing clearly on the terminal/repo-first axes (Terminal-Bench 2.1 81.0 vs Qwen3.8 Max 89.6; DeepSWE 49.3 vs Kimi K3 70.6; NL2Repo 50.0 vs 63.2). That shape says research/office/browser agent loops rather than terminal-first coding agents, but the full release improves on the Preview on nearly every row (BFCL 77.0→78.4, CyberGym 86.5→86.8, AutomationBench 53.8→54.7, Terminal-Bench 78.3→81.0) and doubles context to 512K. Caveats: 744B means frontier memory (FP8 sibling ≈745 GB is the practical path), the custom glm_moe_dsa architecture needs SGLang/vLLM GLM-5.2 recipes, and every number is lab-reported with no independent re-run yet. Included despite 0 downloads because launch-day low counts are not a disqualifier for an established lab (JetBrains Mellum precedent).
See also
- welcome
- Source: https://huggingface.co/internlm/Atria-Dawn
- Paper: https://arxiv.org/abs/2609.15818
- Preview (already tracked): internlm—Atria-Dawn-Preview
- Base model: zai-org—GLM-5.2
- FP8 sibling: https://huggingface.co/internlm/Atria-Dawn-FP8 · Official table: https://atria-asi.ai/
- Related: deepseek-ai—DeepSeek-V4-Pro-0813 · moonshotai—Kimi-K3 · Qwen—Qwen3.8-2.4T-A95B