GLM-5.2 FP8 Uncensored — Colibri int4

This is a Colibri-compatible int4 conversion of zandenAI/GLM-5.2-FP8-Uncensored.

It was created for JustVugg/colibri: a C inference engine that keeps the dense part in RAM and streams routed MoE experts from local storage.

This is an independent conversion, not an official Z.ai, NVIDIA, or Hugging Face release.

Important compatibility note

This repository is not a standard Transformers, vLLM, SGLang, AWQ, or GGUF checkpoint. The tensors use Colibri's custom safetensors container and quantization layout. Use it with Colibri; do not expect AutoModelForCausalLM.from_pretrained() to load it.

Quantization layout

  • Routed expert and ordinary linear weights: Colibri row-wise int4.
  • Token embedding and LM head: int8.
  • Norms and routing weights: FP32.
  • DSA indexer and MTP head: int8.

The source FP8 tensors are dequantized, then requantized into Colibri's format. The original FP8 checkpoint is not included here.

Files

  • out-*.safetensors — main model shards.
  • out-idx-*.safetensors — DSA / IndexShare indexer weights.
  • out-mtp-*.safetensors — MTP head for speculative decoding.
  • config.json, tokenizer files, and generation config — model metadata.

All three groups are required for the complete conversion.

Usage

Follow the build and runtime instructions in the Colibri README. Download the complete repository to fast local storage; Colibri streams expert shards from disk during inference.

License

The source model repository declares the MIT License. This conversion is released under MIT on the same basis; please also review and comply with the upstream model terms.

Credits

Downloads last month
1,028
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for annelo/GLM-5.2-FP8-Uncensored-Colibri-Int4

Quantized
(2)
this model