Tuesday, September 8, 2026

mlx-vlm 0.7.0 ships for Apple Silicon VLMs.

After Monday Issue 11 close, Blaizzy mlx-vlm 0.7.0 landed on GitHub and PyPI on September 7 afternoon ET — a major multimodal MLX release (MoE offloading, redesigned APC, Qwen3.8-Flash-Next MTP + continuous batching + APC disk caching, DeepSeek-V4 Flash Vision EXP / DSpark, LLaVA-OneVision, GLM-5.3, and more), requiring mlx>=0.32.2. This is mlx-vlm, not an mlx-lm bump (text-only mlx-lm still 0.31.3). llama.cpp nightlies continue through b10858 (Vulkan unary+mul — not Metal-primary). Enterprise, frontier flagships, and standing OSS serving engines were quiet this window.

Window after Issue 11 close (~2026-09-07 morning ET) through 2026-09-08 morning. Issue 11 already covered Gemini Enterprise Projects GA, SGLang 0.5.19, Ollama v0.34.0-rc1, LLaDA2.2-mini. Primary sources only. Vendor evals flagged.