Model releases · 11 August 2026
Z.ai's current flagship pairs a million-token context with a sparse attention scheme designed specifically to make that length affordable — and ships it with no regional restrictions attached.
| Released by | Z.ai |
|---|---|
| Published | June 2026, last updated July 2026 |
| Architecture | Mixture-of-Experts with sparse attention (GlmMoeDsaForCausalLM) |
| Layers / hidden size | 78 / 6144 |
| Experts | 256 routed, top-8 per token, plus 1 shared |
| Context length | 1,048,576 tokens (1M) |
| Vocabulary | 154,880 tokens |
| Licence | MIT |
| Weight format | safetensors — 1,506.7 GB across 282 shards |
| Also published | GLM-5.2-FP8 quantised build |
| Hosted at | zai-org/GLM-5.2 on Hugging Face |
GLM-5.2 is Z.ai's flagship model for long-horizon tasks — work that runs across many steps rather than single exchanges. Z.ai frames it as a substantial step over GLM-5.1 and, for the first time in the family, one that delivers on what they call a "solid" 1M-token context: not merely accepting a million tokens, but sustaining useful work across them.
The architectural claim is specific and measurable. IndexShare reuses the same indexer across every four sparse attention layers, which Z.ai reports as a 2.9× reduction in per-token FLOPs at a 1M-token context length. They also report improving the multi-token-prediction layer used for speculative decoding, raising acceptance length by up to 20%.
The model type string, glm_moe_dsa, reflects the sparse attention design rather than a conventional dense-attention MoE.
Z.ai's own release note calls it "Pure Open: an MIT open-source license — no regional limits, technical access without borders." That is unusually direct language, and it puts GLM-5.2 alongside Motif 3 and DeepSeek V4 Flash in the group of very large models with no acceptable-use appendix attached.
The main repository is 1,506.7 GB across 282 safetensors shards. Z.ai also publishes GLM-5.2-FP8, a quantised build that has drawn more than two million downloads on its own — a reasonable indication that most real deployments start there rather than with the full-precision weights.
The wider family on the same account includes GLM-5, GLM-5.1, GLM-4.7 and GLM-4.7-Flash, plus specialised GLM-OCR and GLM-ASR models, all MIT.
| Runtime | Supported | Notes |
|---|---|---|
| transformers | Yes | Standard configuration published by the lab. |
| Official quantised build | Yes | GLM-5.2-FP8 is published by Z.ai alongside the full-precision weights. |
| llama.cpp | No | glm_moe_dsa sparse attention has no announced upstream support. |
| MLX | No | No conversion published by the lab. |
MIT, which Z.ai describes as carrying no regional limits and no additional access restrictions.
A sparse attention design that reuses the same indexer across every four attention layers. Z.ai reports it reduces per-token FLOPs by 2.9 times at a 1M-token context length.
Yes. GLM-5.2-FP8 is published by Z.ai and has more downloads than the full-precision repository.
1,506.7 GB across 282 safetensors shards for the full-precision weights. The FP8 build is substantially smaller.
Not as of publication. The glm_moe_dsa sparse attention architecture has no announced upstream support.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026