Colibri
Colibri refers to the implementation or methodology for running large-scale language models, specifically the glm-52 architecture, on consumer-grade hardware without dedicated GPU acceleration.
Key Characteristics
- Model Architecture: Utilizes glm-52 with a 744 billion parameter count.
- Inference Method: Employs mixture-of-experts (MoE) techniques to optimize memory usage.
- Hardware Requirements: Capable of running on a single machine with approximately 25GB of RAM.
- Acceleration: Performs inference entirely in RAM, requiring no GPU.
Recent Developments
- 2026-07-22: Analysis of Colibri’s capability to run 744B MoE LLMs on consumer laptops, highlighting the feasibility of executing models that theoretically exceed standard memory limits. See Colibri: Unlocking 744B MoE LLMs for Consumer-Grade Laptops.
- 2026-07-14: Documentation of local inference capabilities for GLM-5.2 (744B) via the Colibri method. See [[lab-notes/2026-07-14-Colibri-Local-GLM-5.2-744B-RAM-Inference-wit