InternLM Atria Dawn Preview
One-line: Shanghai AI Lab’s new-generation agentic model, built on the 744B MoE GLM-5.2 foundation — best-in-row BFCL v4 77.0, AutomationBench 53.8, CyberGym 86.5, BrowseComp 92.5. ⚠ large — server-class only. ⚠ lab-reported numbers only so far; surfaced via aggregator (invisible to the HF streams: no pipeline_tag, 546 downloads).
Released 2026-09-11 (llm-stats lists Sep 10), preview + arXiv 2609.15818. Missed by Streams A/B/C because
pipeline_tagis empty and downloads (546) sit far below Stream B’s 1000 cutoff; caught via the llm-stats aggregator + per-org sweep. High engagement for a fresh lab drop (125 likes / 546 dl).
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 | 59.6 | lab |
| SWE-bench Multilingual | — | not published |
| Terminal-Bench 2.1 | 78.3 | lab |
| MLE-bench Lite | 86.2 | lab |
| SkillsBench | 66.4 | lab |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL v4 | 77.0 (best in row) | lab |
| τ-Bench | — | not published |
| τ³-Bench Banking | 41.2 | lab |
| ToolACE | — | not published |
| GAIA | — | not published |
| AutomationBench | 53.8 (best in row) | lab |
| CyberGym | 86.5 (best in row) | lab |
| Workspace-Bench | 65.0 | lab |
| BrowseComp | 92.5 (best in row) | lab |
| DeepSearchQA | 96.0 (best in row) | lab |
| DeepResearch Bench II | 51.1 | lab |
| JobBench | 50.3 | lab |
| GDPval | 1583 | lab |
All numbers lab-reported from the model card’s comparison table (vs DeepSeek V4 Pro 0813, Kimi K3, Qwen3.8 Max, GLM-5.3, GPT-5.6 sol, Claude Opus 5). No independent source (Artificial Analysis / BenchLM) found yet — treat as unverified.
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published |
| GPQA Diamond | — | not published |
| MATH | — | not published |
| AIME 2025 | — | not published |
Efficiency
- Active params: not disclosed (744B-total MoE; config
glm_moe_dsa, 78 layers, hidden 6144, 64 KV heads — DSA = dense/sparse attention variant of the GLM-5.2 family) - Context length: 256K shipped (config layout allows 1M
max_position_embeddings) - VRAM (fp16 rough): ~1.5 TB — ⚠ large, multi-GPU/server-class only; official FP8 variant (
internlm/Atria-Dawn-Preview-FP8) ≈ 745 GB; not single-80GB-GPU hardware - Note: custom architecture
GlmMoeDsaForCausalLM— run via SGLang ≥0.5.13.post1 or vLLM ≥0.23.0 (GLM-5.2 recipes); text-input only (not multimodal)
What makes it notable
Atria Dawn Preview is the strongest open-weight agentic release of the week and the only one the HF streams missed: Shanghai AI Lab’s answer to DeepSeek V4 Pro / Kimi K3 / GLM-5.3, it takes the best-in-row slot on BFCL v4 (77.0 vs V4 Pro 71.4, GLM-5.3 74.1), AutomationBench (53.8), CyberGym (86.5 vs V4 Pro 83.3), BrowseComp (92.5) and Workspace-Bench (65.0), while trailing on Terminal-Bench 2.1 (78.3 vs Qwen3.8 Max 89.3) — so it skews toward research/office/browser agent loops more than terminal-first coding. It ships first-party Codex / Claude Code / Kimi Code integration docs and a hosted API (atria-asi.ai) — unusually agent-native for a lab drop. Caveats: all numbers lab-reported, preview status, active-params undisclosed, and the custom architecture means local serving rides on SGLang/vLLM GLM-5.2 recipes rather than stock transformers.
See also
- welcome
- Source: https://huggingface.co/internlm/Atria-Dawn-Preview
- Paper: https://arxiv.org/abs/2609.15818
- Base model: zai-org—GLM-5.2
- Related: deepseek-ai—DeepSeek-V4.1-Flash