GLM-5.2-Hebrus-GGUF

GLM-5.2 — Hebrus ExpertMajor v2

This repository publishes the qualified GLM-5.2 artifact for andreaborio/hebrus, a source-built Apple Metal inference engine with bounded SSD expert streaming.

The routed MoE bank is stored once in an embedded, checksummed ds4.expert_major.v2 store. The DS4 token in the existing filename and the serialized tensor identifier is a compatibility field; the runtime and Hugging Face library name are Hebrus.

This is a Hebrus-native GGUF extension, not a portable GGUF. Generic llama.cpp, MLX, Ollama, and hosted Hugging Face inference cannot execute it.

Qualified artifact

The legacy chunked diagnostic file and selected-hotlist sidecar under legacy/ remain public for historical reproducibility only. Current Hebrus rejects those layouts and does not use them as fallbacks.

Quick start

Requirements:

• an Apple Silicon Mac with at least 64 GiB unified memory;
• Xcode Command Line Tools;
• more than 262 GB of free local SSD space for the exact model file;
• the official Hugging Face CLI.

git clone https://github.com/andreaborio/hebrus.git
cd hebrus
git checkout v0.3.0

xcode-select --install
python3 -m pip install -U huggingface_hub hf_xet
make -j

./download_model.sh glm-v2

./hebrus \
  -m gguf/GLM-5.2-DS4-ExpertMajor-v2-Q2_K.gguf \
  --ctx 8192 \
  -p "Explain bounded SSD expert streaming in one paragraph."

The downloader pins ds4-v0.2.0 and verifies the complete byte count and SHA-256. The repository is public and ungated, so downloading does not require a Hugging Face token.

Normal AUTO startup resolves to the qualified SSD path. A resident request is rejected even on a larger host. Do not add a sidecar, preload policy, manual cache size, model-admission bypass, or backend flag.

Compatibility and limits

• Apple Silicon Metal with SSD streaming is the only qualified production path for this artifact.
• The 64 GiB floor is not a promise that every context or concurrent workload fits. Start at 8K; 32K has recorded evidence but is slow to test.
• No 65K or 100K GLM support claim is made by this release.
• CPU is reference/debug only. CUDA and ROCm sources are absent, and distributed inference is retired.
• ExpertMajor v1, sidecars, the legacy chunked layout, canonical GGUF inference, and admission bypasses are rejected.
• MTP/speculative decoding is not part of the qualified GLM startup path.
• Performance depends on the exact artifact, runtime revision, context, prompt, storage state, memory pressure, routing locality, and thermals.
• inference: false is deliberate: the Hub hosted inference widget does not support this runtime-specific artifact.

See the authoritative runtime support contract and the scoped 32K evidence.

License and attribution

The upstream zai-org/GLM-5.2 repository declares the MIT license. Z.ai retains the base-model copyright. Historical runtime work credits antirez/DS4; the embedded ExpertMajor v2 release and current runtime are maintained by Hebrus.