Atlas

Models

← All models

GLM 4.6

Zhipu AI (Z.ai) · 2025-09-30 · 356.8B parameters

GLM-4.6 is Zhipu AI's late September 2025 open-weight MoE language model marketed as 355 billion total parameters with 32 billion active, featuring a 200K-token input context and 128K max output. It improves on GLM-4.5 with enhanced real-world coding performance, longer context handling, stronger reasoning, and better agentic tool use during inference. Released under the MIT license, it integrates directly with popular coding agent environments.

Benchmark scores

BenchmarkScore
AA-LCR54.3
AA-Omniscience Index-31.6
AIME 2024 (avg@8)94.6
AIME 2024-202592.7
AIME 202544.3
AIME 2025 (avg@8)90.8
ALE-Bench838
Apex (MathArena)0.5
APEX-Agents4.0
Arena Code WebDev — Brand & Marketing1348
Arena Code WebDev — Consumer Product1355
Arena Code WebDev — Content Creation Tools1339
Arena Code WebDev — Data & Analytics1342
Arena Code WebDev — Gaming1332
Arena Code WebDev — HTML1349
Arena Code WebDev — Reference-Based Design1340
Arena Code WebDev — Simulations1323
Arena Text — Business, Management & Financial Ops (No Style Control)1418
Arena Text — Chinese (No Style Control)1500
Arena Text — Creative Writing (No Style Control)1413
Arena Text — English (No Style Control)1448
Arena Text — Entertainment, Sports & Media (No Style Control)1409
Arena Text — Exclude Ties (No Style Control)1438
Arena Text — Exclude Ties (Style Controlled)1417
Arena Text — Expert (No Style Control)1434
Arena Text — French (No Style Control)1462
Arena Text — French (Style Controlled)1447
Arena Text — German (No Style Control)1450
Arena Text — Hard Prompts (No Style Control)1441
Arena Text — Hard Prompts English (No Style Control)1444
Loading Atlas data…