Big AI Models Without a GPU: colibri Reads Experts From Disk

The open-source engine colibri can run mixture-of-experts models with hundreds of billions of parameters on an ordinary computer with no graphics card. It keeps only the part of the model that is needed in RAM and loads the “experts” from disk whenever the model calls for them. In the README example, a chat session uses GLM-5.2, a 744-billion-parameter model at int4 quantization, runs on a CPU alone and is ready after 32 seconds.

The project is written in plain C with no dependencies, and each model family gets its own file. Thirteen engines work at the moment. Ten handle language models, among them the GLM line (Mistral-Large-4 runs on the same engine), Kimi K3, DeepSeek V4 and V4.1 Flash, MiMo-V2.6, Qwen, Inkling and OLMoE. Another engine generates images with Qwen-Image-2.1, and two more, Laya and GLiNER2.5-Decide, produce decisions. A third such model, Clef, runs on the Qwen3.6 engine.

Beyond its own formats, colibri can load published GGUF files directly, with no conversion. Compute can run on the CPU, on any GPU that supports Vulkan, or on NVIDIA hardware via CUDA. The code is open under the Apache-2.0 license, and the current release is v2.1.0.

The repository has 46,859 stars and 5,410 forks. It was created on July 1, 2026, which works out to about 375 new stars a day over 125 days. The latest update is dated November 3.

Other projects also run models locally on a CPU, llama.cpp among them, which Georgi Gerganov started in March 2023. colibri bets on something different: very large MoE models, where only a small slice of the parameters is active for each token. That is why the engine leaves the remaining experts on disk and reads them on demand.