tokenizer

This release introduces significant new features like ETags support, a Websocket extension, and advanced `hx-boost` overrides, alongside numerous bug fixes and improvements to existing functionalities like Morph and SSE handling. Several internal refactors, such as moving the tokenizer and extracting the SSE extension, may require migration steps for existing extensions.

This release introduces model batching support and fixes several issues related to tokenization and desktop application startup behavior. It also includes support for NVIDIA TensorRT Model Optimizer import.

This release introduces model batching support and TensorRT Model Optimizer import for the mlx backend. It also includes several bug fixes related to tokenization and desktop application startup behavior.

This release introduces model batching support and adds NVIDIA TensorRT Model Optimizer import capability. Several minor bugs related to tokenization and desktop app session handling were also resolved.

This release introduces support for NVIDIA TensorRT Model Optimizer import within mlx and fixes an issue related to multi-regex BPE offset handling in the tokenizer. It also includes performance improvements by batching the sampler across multiple sequences in mlxrunner.

This release introduces the new Gemma 4 model family variants (E2B, E4B, 26B, 31B) and enhances tokenizer capabilities with SentencePiece-style BPE support.

This release introduces the new Gemma 4 model family variants (E2B, E4B, 26B, 31B) and enhances tokenizer capabilities with SentencePiece-style BPE support.

This release introduces UI exposure of the server context length and implements OpenClaw onboarding, alongside internal consolidation of the tokenizer.

This patch release focuses on bug fixes for tool calling, generation errors, and tokenization issues across specific model architectures.

This release addresses critical memory leaks within the WebGPU implementation and resolves several related testing and device handling bugs.

This release introduces major performance enhancements with FP8 Reinforcement Learning support and significant VRAM reductions across the board. It also adds support for new models like DeepSeek-OCR and Qwen3-VL, alongside improved Docker integration.