jundot/omlx v0.2.12

on GitHub

Download the DMG that matches your macOS version (sequoia or tahoe).
If you're on an M5 Mac, you must use the macos26-tahoe DMG for M5 Neural Accelerator.

Highlights

• ModelScope Hub integration for users in China with limited HuggingFace access (#229)
• Model settings modal redesign — split into Basic/Advanced two-column layout with experimental features section
• Experimental IndexCache optimization for DSA models (DeepSeek V3.2, GLM-5) — skip redundant indexer computation for faster inference (#214). GitHub, arXiv

New Features

ModelScope model downloader (#229)

• native ModelScope Hub integration for discovering and downloading MLX models
• SDK-based search with mlx-community organization filtering
• download cancellation with post-completion cleanup
• memory-aware model filtering with size display and tight-fit warnings
• --ms-endpoint CLI argument for custom ModelScope endpoint

Model settings modal redesign

• two-column layout: basic settings (left) and advanced settings (right)
• experimental features section with IndexCache toggle and GitHub link
• chat template kwargs now use a two-line layout to prevent overflow
• prevent duplicate enable_thinking / reasoning_effort kwargs

IndexCache for DSA models (#214)

• experimental optimization that skips redundant indexer computation in MoE+MLA layers
• configurable frequency (default: every 4th layer keeps its indexer)
• toggle available in model settings for DSA models only (DeepSeek V3.2, GLM-5)

HuggingFace search filter (#222)

• filter HF model search results to MLX library only, reducing noise from non-MLX models

Integration improvements

• detect install method and show correct CLI command in integrations settings (#225)
• include model context/output limits in OpenCode integration (#231)

Bug Fixes

• fix tool-call markup leaking into reasoning content
• fix tool calls generated inside <think> blocks by small models not being recovered
• fix macOS resource fork files (._*) causing model deletion failures on exFAT/NTFS external drives (#228)
• fix OCR model prompts being overwritten instead of preserved
• fix auto-update requiring a second click after download
• fix index_cache_freq missing from model list API response, causing settings to reset on modal reopen
• fix shutil.rmtree onerror handler losing original traceback

New Contributors

• @ChiangJingYing made their first contribution in #228
• @bbulee made their first contribution in #229