Atlas

Models

← All models

Qwen3.8 27B

Alibaba · 2026-08-14 · 27.8B parameters

Qwen3.8 27B is Alibaba's dense, post-trained vision-language checkpoint for coding and agentic work, accepting images and videos. It supports optional thinking with low, medium, and xhigh effort, and a native 262,144-token context extendable to 1,000,000 with the documented scaling configuration. The complete BF16 checkpoint contains 27,781,427,952 parameters; its tokenizer assigns 248,077 distinct IDs, while the padded embedding configuration has 248,320 rows.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo1398
AA-Briefcase v1.1 Elo1393
AA-LCR v1.179.7
AA-Omniscience Index-8.0
AA-Omniscience v1.1 Accuracy18.3
AA-Omniscience v1.1 Hallucination Rate18.0
AA-Omniscience v1.1 Index-6.4
AgentVQA50.9
AgentVQA Egocentric Videos69.0
AgentVQA Games61.5
AgentVQA Robotics42.5
AgentVQA Spatial Planning51.5
AgentVQA Web Agents30.0
ARC-AGI-1 Semi-Private87.5
ARC-AGI-242.4
Arena Alignment — Deceptive Completion19.3
Arena Alignment — False Attribution9.9
Arena Alignment — Unauthorized Action2.3
Arena Alignment Index73.6
Arena Code WebDev — Brand & Marketing1608
Arena Code WebDev — Consumer Product1559
Arena Code WebDev — Content Creation Tools1556
Arena Code WebDev — Data & Analytics1551
Arena Code WebDev — Frontend1598
Arena Code WebDev — Fullstack1615
Arena Code WebDev — Gaming1634
Arena Code WebDev — HTML1569
Arena Code WebDev — React1600
Arena Code WebDev — Reference-Based Design1597
Arena Code WebDev — Simulations1598
Loading Atlas data…