GLM 5.2

GLM-5.2 is a text-generation model designed for long-horizon tasks, supporting a solid 1M-token context with advanced reasoning and coding capabilities.

Key Strengths

• 1M-token context that stably sustains long-horizon work.
• IndexShare architecture reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at 1M context length.
• Improved MTP layer for speculative decoding increases acceptance length by up to 20%.
• Three thinking modes: Non-thinking, Thinking High, and Thinking Max, enabling flexible effort trade-offs.
• Mixture-of-Experts with 744B total parameters and 40B active parameters, released under MIT license.

Benchmark Performance

GLM-5.2 achieves competitive scores across reasoning, coding, and agentic benchmarks. Notable results include HLE 40.5, HLE with Tools 54.7, AIME 2026 99.2, SWE-bench Pro 62.1, and Terminal Bench 2.1 81.0.

Quantization and Hardware

• Dynamic 1-bit (UD-IQ1_M) reaches ~76.2% top-1 accuracy while being 86% smaller; dynamic 2-bit ~82% accuracy at 84% smaller; dynamic 4-bit and 5-bit are described as “mostly lossless.”
• Quantized model sizes range from 223 GB (1-bit) to 810 GB (8-bit).
• 2-bit quant fits on a 256 GB unified memory Mac; works with MoE offloading on 1×24 GB GPU + 256 GB RAM.

Full documentation and analysis and technical report are available.