Skip to content
Model Pareto

Atria Dawn Preview

Shanghai AI Lab (Intern)Open weightshuggingface.co ↗
Compare with nearest rivalsCompare with…
Input
—
Output
—
Speed
~81tok/sest.
Context
262Ktokens

Agentic model from Shanghai AI Lab built on GLM-5.2 (744B MoE). No API price found. Card also reports self-run Terminal-Bench 2.1, AutomationBench, GDPval and tau3-Banking, not mapped (our ids hold AA / Terminal-Bench 4.0 runs).

Rankings

Self-hostingEstimated

How we estimate →

What it would cost to run these open weights yourself on rented GPUs. No API sells this model, so this is its price on the chart.

Hardware
8× B200 180GB
FP8 weights
Throughput
~18,596 tok/s
many requests batched
Price
$0.30–2.69 /M tok
busy → light use
On your own machine
Mac Studio M5 Ultra 512GB (int4, ~42 tok/s single-stream)
single consumer GPU or Mac
Assumptions (5)
  • FP8 weights (744B params, 40B active per token) + 25% KV-cache headroom ≈ 930 GB
  • 8× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
  • ~18,596 output tok/s aggregate at batch 532 (bandwidth-bound); MoE compute scales with active params
  • Blended 3:1 input:output; prefill ~212,625 tok/s
  • Low = 75% utilization at the low GPU price; high = 20% at the high price

Benchmark scores

~ italic, dashed = no published score yet, estimated from related benchmarks.

Coding

SWE-bench VerifiedEstimated
~#3 of 14 (estimated)~81.9%
  1. #1Claude Opus 596.0%
  2. #2Claude Sonnet 585.2%
  3. ~#3Atria Dawn PreviewEstimated81.9%
  4. #3Gemini 3.1 Pro80.6%
No published score — estimated from related benchmarks
DeepSWE v1.1Estimated
~#27 of 31 (estimated)~54.2%
  1. #1DeepSeek V4.1 Flash74.2%
  2. #2Grok 4.772.6%
  3. #26Nex-N2.5-Pro55.8%
  4. ~#27Atria Dawn PreviewEstimated54.2%
  5. #27Qwen3.8 Max (0902)51.0%
No published score — estimated from related benchmarks
LiveCodeBenchEstimated
~#6 of 20 (estimated)~81.3%
  1. #1Qwen3.8-Omni-Flash92.6%
  2. #2Qwen3.8 Flash-Next91.9%
  3. #5gpt-oss-120b87.8%
  4. ~#6Atria Dawn PreviewEstimated81.3%
  5. #6Gemma 4 31B80.0%
No published score — estimated from related benchmarks
SciCodeEstimated
~#12 of 96 (estimated)~57.3%
  1. #1Claude Opus 5.566.9%
  2. #2Claude Fable 5.163.1%
  3. #11Grok 4.757.4%
  4. ~#12Atria Dawn PreviewEstimated57.3%
  5. #12Gemini 3.8 Flash56.6%
No published score — estimated from related benchmarks
AA Coding Agent IndexEstimated
~#7 of 15 (estimated)~55.1
  1. #1Claude Opus 5.566.0
  2. #2Claude Fable 5.162.2
  3. #6Grok 4.756.3
  4. ~#7Atria Dawn PreviewEstimated55.1
  5. #7Muse Spark 1.354.3
No published score — estimated from related benchmarks

Agents

τ²-bench (Telecom)Estimated
~#26 of 116 (estimated)~82.5%
  1. #1Step 3.7 Flash98.5%
  2. #2Gemini 3.1 Pro95.6%
  3. #24Nemotron 3 Ultra83.3%
  4. ~#26Atria Dawn PreviewEstimated82.5%
  5. #26Nex-N2-Pro81.6%
No published score — estimated from related benchmarks
OSWorld-VerifiedEstimated
~#3 of 12 (estimated)~82.6%
  1. #1Kimi K384.8%
  2. #2Qwen3.8 27B84.3%
  3. ~#3Atria Dawn PreviewEstimated82.6%
  4. #3Nex-N2.5-Pro82.2%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~#3 of 5 (estimated)~57.9%
  1. #1GLM-5.3 Flash59.1%
  2. #2Kimi K358.3%
  3. ~#3Atria Dawn PreviewEstimated57.9%
  4. #3GPT-5.6 Terra50.2%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Estimated
~#18 of 87 (estimated)~1458 Elo
  1. #1Claude Opus 5.51846
  2. #2Claude Fable 5.11735
  3. #17GPT-6 Sol1487
  4. ~#18Atria Dawn PreviewEstimated1458
  5. #18Claude Sonnet 51449
No published score — estimated from related benchmarks