macMLX questions, answered

macMLX supports Apple Silicon Macs running macOS 14 or later. Follow the project's current installation and Gatekeeper guidance for the unsigned app; do not disable system-wide security protections solely to open macMLX.

Not for the default released engine, app, or CLI inference path. Optional compatibility engines may use other runtimes, so macMLX does not claim that Python is absent everywhere.

Choose an MLX checkpoint whose exact architecture is supported. For capacity, subtract macOS and app headroom from physical unified memory, then budget weights, KV cache, activations, and any concurrent models; if the total approaches the remainder, choose a smaller or more strongly quantized checkpoint.

v0.9.0 supports scoped OpenAI endpoint families, including audio transcription and speech, plus Anthropic Messages and five selected Ollama endpoints, not provider-wide drop-in compatibility. Released controls include multi-turn tool loops, structured output, logprobs, XTC, per-request adapters, and KV-cache quantization; unsupported schemas, parameter combinations, and audio containers return explicit errors.

The default inference engine and HTTP server run locally on the Mac. Network access is still used when you download models, query remote registries, or configure external tools.

Yes. The v0.9.0 model library detects exactly 14 VLM model_type families. That count is a family-level discovery signal, not universal checkpoint support: verify the exact checkpoint, image processor, weight variant, quantization, memory budget, and serving path before relying on generation.

DeepSeek V3.2 still lacks a real-checkpoint smoke test and FP8 dequantization, so its Swift component parity is not an end-to-end inference claim. Track G adds checkpoint-tested families, but those measured results are checkpoint-specific and do not establish universal inference support for other large MoE families.

v0.9.0 has released eligibility-gated batching, longest-prefix reuse, XTC, and KV-cache quantization. Paged KV, block sharing, copy-on-write branching, and the adaptive memory guard remain planned, as do user-facing top-k, min-p, presence, frequency, and repetition penalties plus a per-request seed. Multi-token prediction is detectable but deliberately not wired into decoding.