AI Potluck
Model components / Base / pretrained models

GLM-5.2

Z.ai

Z.ai's flagship open-weights frontier model (mid-June 2026, successor to GLM-5.1), built for long-horizon agentic and coding tasks. A sparse MoE (~753B total params) with a 1M-token context, introducing 'IndexShare' sparse attention (~2.9x FLOP cut at 1M context) and improved speculative decoding. Positioned as a leading open-weights coding model.

MIT-licensed open weights (HF model card + LICENSE, (c) Zhipu AI 2026). ~753B-param MoE, 1M context. Consolidated on 2026-07-29 from glm-4, glm-4-6, glm-4-9b, glm-5-1; openness follows glm-5-2, the release that currently governs. See docs/reference/identity.md. GLM-5.2 has pulled away sharply since July and now supplies most of the family's downloads, so the adoption band, which was set on the GLM-4 generation, is a level low and flagged for review. Verified 2026-08-13 via the GLM-5.2 LICENSE file, its HF model card and the zai-org/GLM-5 repository.

Openness

3 high confidence
3.0
weights
open(MIT)
data
closed
code
open(inference)
license
MIT(OSI)

MIT weights on the current release, data closed, no training pipeline, so 3. Z.ai moved to MIT partway through the family: GLM-4 and GLM-4-9B carried the GLM-4 License, which requires commercial-use registration and a 'Built with GLM' attribution and scored 2, while GLM-4.6, GLM-5.1 and GLM-5.2 are plain MIT. GLM-5.2 governs as the current release. Capability 5 comes from the GLM-5 line; the GLM-4 line scored 2.

Adoption

4 high confidence
4.0

2,774,206 downloads in the trailing 30 days across the four declared artifacts (zai-org/GLM-4.6 28,016; zai-org/glm-4-9b 11,688; zai-org/GLM-5.1 72,174; zai-org/GLM-5.2 2,662,328), which bands at level 4 (1M-10M) on the model adoption scale. GLM-5.2 alone supplies 96% of that total, so the current release carries the family rather than a generation-old checkpoint. The older THUDM/glm-4-9b repository now redirects to zai-org.

Capability

5 high confidence
5.0

SWE-Bench Pro 58.4 (reported #1 at release, edging GPT-5.4 57.7 and Claude Opus 4.6 57.3); Terminal-Bench 2.0 63.5. Frontier-tier agentic coding for an open-weights model.

Verified 2026-08-13