b10437
ggml-org/llama.cppb10437Aug 15, 2026by github-actions[bot]same day as b10436
What changed since b10436
Adds MiniMax-Text-01 and MiniMax-M1 model support with recurrent-state and logits-mask optimizations.
New features
- Added support for the MiniMaxText01ForCausalLM architecture.
- Added support for MiniMaxM1ForCausalLM, reusing the same conversion path as MiniMax-Text-01.
- Introduced a generic logits-mask graph input to suppress zero-valued embedding tokens during sampling.
- Added a Jinja chat template for MiniMax-M1.
Improvements
- Replaced hardcoded model checks with the hparams.is_recr() helper for recurrent models.
- Switched recurrent state handling to the shared build_rs() path.
- Removed redundant state transpose and ggml_cont() operations from the MiniMax-Text-01 graph.
- Pre-permuted diag_decay dimensions and used token positions for decay tensor computation to cut per-step work.
- Conversion now yields while modifying tensors and suppresses zero-embedding tokens for MiniMax-Text-01.
Bug fixes
- Fixed token sampling disruption caused by zero logits from MiniMax-Text-01 padding embeddings.
- Corrected MiniMax-01 related test failures and silenced associated Python lint errors.
Upgrade notes
No configuration changes are required for existing models; the new architectures are picked up automatically by the converter and loader. If you run MiniMax-Text-01 or M1, re-convert GGUF files with the updated converter so zero-valued embedding tokens are suppressed. The earlier logits-mask workaround was superseded during the PR, so no manual masking flag is needed.
Analyzed Oct 24, 2026
Upgrading?
Release notes summary
Adds support for MiniMax Text and M1 models with specific optimizations including logits masking and recurrent state management.
Key Highlights
- Added MiniMaxText01ForCausalLM support
- Added MiniMaxM1ForCausalLM support
- Optimized recurrent state management
- Removed unnecessary tensor operations
- Added jinja template for MiniMax-M1
Release Notes (excerpt)
<details open> model : add support for MiniMaxText01ForCausalLM and MiniMaxM1ForCausalLM (#27018) * llama : support for MiniMax-Text-01 model * chore : renames to match the other MiniMax models * model : add logits mask as MiniMax-Text-01 embeddings tensor has zero-valued embeddings for tokens >= 200032 that produce zero logits disrupting the token sampling process * llama : replace hardcoded conditions with hparams.is_recr() * model : used build_rs() for recurrent state management…
Published by ggml-org on GitHub.
Read the full release notesRelated Repositories
:electron: Build cross-platform desktop apps with JavaScript, HTML, and CSS
Godot Engine – Multi-platform 2D and 3D game engine
ClickHouse® is a real-time analytics database management system
Self-hosted game stream host for Moonlight.
DuckDB is a high-performance analytical database system