Llama.cpp Adds MiMo V2.5 Model Support (Sparse MoE 310B)
Patch your llama.cpp to the branch with PR #22493, compile with default flags, and load the MiMo V2.5 GGUF to gain 1 M‑token context and multimodal support.
Patch your local llama.cpp to the PR #22493 branch, compile, and load the MiMo V2.5 GGUF to enable 1 M‑token multimodal inference.
Summary
A new pull request (PR #22493) to the llama.cpp repository adds support for the Xiaomi MiMo V2.5 model, a sparse Mixture‑of‑Experts architecture with 310 B total parameters and 15 B activated per inference. The model supports up to 1 M token context and handles text, image, video, and audio modalities. Its vision encoder is a 729 M‑parameter ViT with 28 layers (24 SWA + 4 full), while the audio encoder is a 261 M‑parameter Audio Transformer with 24 layers (12 SWA + 12 full). The Multi‑Token Prediction head contains 329 M parameters across three layers, enabling speculative decoding. The PR makes the model available on Hugging Face at https://huggingface.co/XiaomiMiMo/MiMo‑V2.5 and includes a screenshot of the pull request on the ggml‑org/llama.cpp GitHub page. Developers can compile the updated llama.cpp branch and load the MiMo V2.5 GGUF files to leverage its 1 M‑token context and multimodal capabilities. The sparse MoE design allows efficient inference with only 15 B active parameters, while the MTP head provides faster generation.
Key changes
- Added MiMo V2.5 model support via PR #22493.
- Model is sparse MoE with 310 B total, 15 B activated parameters.
- Supports up to 1 M token context and multimodal (text, image, video, audio).
- Vision encoder: 729 M‑parameter ViT (28 layers: 24 SWA + 4 full).
- Audio encoder: 261 M‑parameter Audio Transformer (24 layers: 12 SWA + 12 full).
- MTP head: 329 M parameters, 3 layers.