GLM 5.2 (IQ2_XXS)
antirez's routed 2-bit community GGUF of Z.ai's GLM 5.2 — a 754B-parameter Mixture-of-Experts (256 experts, 8 active per token, ~40B active) with a 1M-token context and DeepSeek-style sparse attention. The routed experts are crushed to IQ2_XXS (block 78 in Q2_K) while attention, the shared expert, embeddings, and the output head stay at Q8-class precision, landing the file at ~197 GB. Only ~20 GB of that is non-routed weight that must stay resident: DwarfStar streams the remaining ~177 GB of experts from SSD, so it runs on a 128 GB machine — but it is a very large download and a slow first turn. Built for antirez's `ds4` engine; these quants do not load on stock llama.cpp.
At a glance
- Parameters
- 754B
- Quantization
- n/a
- Context window
- 1049K tokens
- Approx. size
- 211.1 GB
- Engines
- n/a
- License
- MIT (unclassified)
- Version
- v1.0.0 , released 2026-07-25
- Family
- glm
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
1- sampling.topP
0.95- sampling.maxTokens
16384- reasoning.thinkingBudget
4096- reasoning.enableThinking
true
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-condensed
- mcp.compact-tool-schemas
- fabrication.detect-past-tense-no-tools
- turn.preamble-folding
Sources
- Upstream
- https://huggingface.co/zai-org/GLM-5.2
- Manifest
- View on GitHub
Tags
- glm
- reasoning
- moe
- tools
- 2-bit
- iq2
- long-context
- large