v1.0.11

Release Summary

1. Updated README.md to streamline content and enhance clarity for easier understanding.
2. Merged Pull Request #27 (contributed by @koush) to remove an unused dependency, reducing potential bloat and improving maintainability.

v1.0.10

Summary of Changes

• Updated to the latest versions of mlx-lm and mlx-vlm to ensure compatibility and access to recent improvements.
• Optimized default image resolution from 512 to 448 for improved performance.

v1.0.9

Summary of Changes

• Implemented OpenAI-Standard Streaming with Tool Calls and Thinking Parser: Fully integrated streaming support for tool calls, including a custom thinking parser, in compliance with OpenAI’s latest API standards. Implementation by @tienthanh214 (#25).
• Logging System Migration: Replaced the standard logging module with loguru to improve readability, flexibility, and ease of debugging.
• Demo Video Update: Updated the demo video in README.md to reflect the latest features and behavior of the current codebase.

v1.0.8

Craft a concise and informative description for PyPI package. The description highlights the key features and purpose of mlx-openai-server package while being clear and engaging.

v1.0.7

Summary of Changes

• Temporarily removed metrics.py pending a better implementation.
• Aligned all Pydantic model fields with the OpenAI schema for consistency.
• Updated finish_reason from "function_call" to "tool_calls" in streaming responses involving tool usage.
• Fixed parsing of tool calls for Qwen3 models.
• Resolved issues with VLM models when handling non-streaming text requests.
• Added torchvision to setup.py — required by some VLM models for image processing.

v1.0.6

Summary

This new release tag to mark the latest stable version of the codebase. Key updates included in this release:

• New Feature: Introduced the /v1/models endpoint for monitoring model serving status.
• Updates: Synced with the latest versions of mlx_vlm and mlx_lm for up-to-date performance and compatibility.
• Bug Fix: Fixed a text extraction issue when processing chunks.
• Enhancement: Refined the resource cleanup logic for improved efficiency and stability.

v1.0.5

BREAKING CHANGE: Rename package and CLI from mlx-server to mlx-openai-server

Summary

This PR introduces a breaking change by renaming the package and CLI from mlx-server to mlx-openai-server to resolve PyPI naming conflicts and improve compatibility.

Changes

• Renamed the package in setup.py from mlx-server to mlx-openai-server
• Updated all CLI references from mlx-server to mlx-openai-server
• Updated the repository/package name in the README and all usage instructions
• Added a "Breaking Change" notice to the README

Impact

Breaking change:
All users must update their scripts, CLI commands, and installation instructions to use mlx-openai-server instead of mlx-server.

Motivation

The original name mlx-server was too similar to existing projects on PyPI, causing upload errors. This change ensures uniqueness and future compatibility.

Migration

• Replace all usage of mlx-server with mlx-openai-server in your scripts and CLI commands.
• Update any installation instructions to use the new package name.

v1.0.4

Automating Package Releases with GitHub Action

v1.0.3

Key Enhancements in This Release:

• Function Calling Support for Qwen3: Added support for function calling with the Qwen3 model, the latest addition to Alibaba Model Studio’s Qwen family of large language models.
• Embeddings Endpoint: Introduced the v1/embeddings endpoint for both LM and VLM, enabling embedding generation.
• Improved OpenAI Compatibility: Refactored the schema to enhance compatibility with OpenAI’s API format.
• RAG Demo Notebook: Added simple_rag_demo.ipynb, showcasing an engaging use case for serving local models using mlx-server-OAI-compat.
• Consistent Error Handling: Standardized all error responses across the codebase using create_error_response for a unified API error format.

v1.0.2

Changes in this release:

• Refactored API schemas and response formats for improved consistency and maintainability.
• Updated chat history handling logic for better performance and reliability.
• Exposed the /v1/embeddings endpoint to support MLX-LM models (text-only).
• Added a new notebook, embeddings_examples, demonstrating how to use the embeddings endpoint via the OpenAI-compatible API.