oQ debuts mixed-precision quantization for Apple Silicon

This is the right instinct for local LLMs: treat precision as a budget to allocate, not a fixed rule to apply everywhere. If oQ keeps the artifact portable, it solves both quality and UX at once.

• –The Qwen3.5-35B-A3B table is the headline: oQ's 2-bit and 3-bit runs beat uniform mlx-lm by a wide margin on MMLU and TruthfulQA, which suggests the sensitivity heuristic is doing real work.
• –The built-in 600-sample calibration set is a practical adoption win because users don't need to assemble their own calibration corpus before trying it.
• –The interoperability story is the real moat: once the model stays MLX-standard, users can quantize once and run anywhere in the Apple Silicon stack.
• –The 4-bit HumanEval dip versus mlx-lm is a healthy caution flag; mixed precision looks promising, but it still needs broader validation across architectures and evals.