Show HN: Colibrì — Running 744B-Parameter GLM 5.2 on Consumer Hardware with 32GB RAM

What happened

HN post 'Show HN: Getting GLM 5.2 running on my slow computer' (730 points) demonstrates Colibrì — an open-source system enabling the 744B-parameter Z.ai GLM 5.2 Mixture-of-Experts model to run on consumer hardware with just 32GB RAM.

Context and impact

GLM 5.2 is one of the most capable open-weight models currently available, competitive with frontier models. Standard deployment requires dozens of GB of VRAM on GPU servers. Colibrì bypasses this by streaming only active expert parameters from disk while keeping inactive MoE experts in storage. The result is extremely slow but functional and accessible to anyone.

Details

• Model: Z.ai GLM 5.2, 744B parameters, Mixture-of-Experts architecture
• Requirements: only 32GB RAM (vs. 100+ GB VRAM for GPU server deployment)
• Performance: ~0.1 tokens per second (slow but functional)
• Method: streaming active experts from disk; inactive experts stay in storage
• Source code: github.com/JustVugg/colibri (open-source, MIT)
• HN score: 730 points (July 9, 2026)