Atlas

Models

← All models

GLM-4.7

Zhipu AI (Z.ai) · 2025-12-22 · 358.3B parameters

GLM-4.7 is Z.ai’s December 22, 2025 open-weight MoE text model for coding, tool use, complex reasoning, and agentic workflows. Z.ai markets the architecture as 355 billion total parameters with 32 billion active per forward pass; the released checkpoint contains exactly 358,337,776,896 trainable parameters and configures 92 transformer layers, 160 routed experts with eight selected per token plus one shared expert, GQA, and partial RoPE over half of each attention head. It supports interleaved, preserved, and per-turn-controllable thinking, has a 202,752-position checkpoint configuration, and is released under the MIT license.

Benchmark scores

BenchmarkScore
AA-LCR64.0
AA-Omniscience Index-34.6
AIME 2024 (avg@8)95.0
AIME 2024-202593.3
AIME 202595.0
AIME 2025 (avg@8)91.7
ALE-Bench1011
APEX (Mercor)6.2
APEX-Agents8.7
APEX-v1-extended51.7
Arena Code WebDev — Brand & Marketing1449
Arena Code WebDev — Consumer Product1435
Arena Code WebDev — Content Creation Tools1410
Arena Code WebDev — Data & Analytics1420
Arena Code WebDev — Gaming1427
Arena Code WebDev — HTML1446
Arena Code WebDev — React1424
Arena Code WebDev — Reference-Based Design1437
Arena Code WebDev — Simulations1403
Arena Text — Business, Management & Financial Ops (No Style Control)1428
Arena Text — Chinese (No Style Control)1484
Arena Text — Creative Writing (No Style Control)1402
Arena Text — English (No Style Control)1454
Arena Text — Entertainment, Sports & Media (No Style Control)1403
Arena Text — Exclude Ties (No Style Control)1432
Arena Text — Exclude Ties (Style Controlled)1442
Arena Text — Expert (No Style Control)1420
Arena Text — French (No Style Control)1427
Arena Text — French (Style Controlled)1443
Arena Text — German (No Style Control)1420
Loading Atlas data…