b10142

ggml-org/llama.cppb10142Jul 27, 2026by github-actions[bot]same day as b10141

What changed since b10141

high risk

Added MiniMax-M3 multimodal model support with vision capabilities and optimized long-context performance.

Breaking changes

  • All GGUFs generated before this change will need to be regenerated.

New features

  • MiniMax-M3 multimodal model support
  • Vision tower support for MiniMax-M3
  • Flash attention for sparse layers
  • Multi-stream support for MSA attention
  • Optimized CPU/GPU ops for long context

Improvements

  • Significant performance improvements for long context
  • Sparse attention with dense fallback
  • Decomposed slow CPU operations into GPU + CPU ops
  • Rewrote indexer operation to be CUDA native

Bug fixes

  • Fixed cache allocation and contiguous memory issue
  • Fixed -fa auto crash

Upgrade notes

Regenerate all GGUF files generated before this change as the model architecture has changed. Note that sparse attention currently falls back to dense attention.

Analyzed Oct 16, 2026

Upgrading?

Release notes summary

Added preliminary support for the MiniMax-M3 multimodal model, including vision capabilities, optimized attention mechanisms, and multi-stream handling.

Key Highlights

  • MiniMax-M3 vision support
  • Sparse attention with dense fallback
  • Optimized flash attention and multi-stream support
  • Significant performance improvements for long context

Release Notes (excerpt)

<details open> mtmd: Add Vision Support for Minimax-M3 (#25113) * Add preliminary MiniMax-M3 support Text-only port that re-uses existing components: MiniMax-M2 style GQA with per-head QK-norm and partial rotary, DeepSeek-V3 style leading-dense and routed/shared experts, and swigluoai activation. Sparse attention is not yet supported (dense fallback); vision tower and MTP heads are dropped. * MiniMax-M3 vision tower (mmproj + clip graph) * Delete m3_vision_ref.py * Update clip.cpp * MSA…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories