MLX-LM – v0.28.3
🔥 MLX LM v0.28.3 is LIVE! 🔥
Heads up, Apple silicon LLM tinkerers – LLaMA-Factory just dropped a massive update for MLX LM! This release is packed with refinements and new features to help you build, train & serve even better models.
Here’s the breakdown:
- Memory Efficiency: State Space Models (SSM) are leaner now. 🙌
MoE Magic: Lots of improvements to Mixture of Experts – LoRA fixes, bailing logic, and* a new LFM2 option!
- Qwen3-VL Support: Visual language model support added with Qwen3-VL (plus a dense version!). 🖼️
- Faster GPT2: Batch processing for GPT-2 just got quicker.
- DWQ Tweaks: Depthwise Quantization refined with temperature adjustments.
- Python 3.9 Love: Qwen3 support now extends to Python 3.9 users!
- Plus: Cleaned up params, simplified I/O, CUDA install fixes, batched SSM masking, gradient accumulation, data parallel eval, Jamba support & LLM Benchmarks! 📊
Dig into the full changelog – there’s a ton here to play with! 🎉