# MLX-LM

Python generation tools for MLX-compatible models on Apple Silicon.

Revision: 2026-09-15.1 · Review: unreviewed

Commissioned by a Nimbalyst maintainer. Documentation synthesis; no independent review or product execution.

## Applicability

Repository revision 872ae88d1fac77350db23c8c04fe8dd372a9e3e8; living documentation separately dated.. Local Apple Silicon execution with MLX; model conversion and server configuration remain operator-owned..

## Claims

- **yes · documented**: Provides local text generation with MLX. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: Model files must be acquired before an offline run.
- **yes · documented**: Targets Apple Silicon through the MLX framework. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: Supported Apple hardware and MLX installation; speed unmeasured.
- **yes · documented**: Loads MLX-compatible model repositories or local conversions. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: Supported architectures and MLX-compatible weight layouts.
- **yes · documented**: Provides conversion to quantized model weights. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: README conversion example uses 4-bit weights; not equivalent to GGUF 4-bit artifacts.
- **yes · documented**: Provides incremental output through stream_generate. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: Python streaming interface; no blanket HTTP API parity claim.
- **yes · documented**: Provides prompt caching for repeated long prompts. [MLX-LM documentation](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/README.md) Scope: mlx_lm.cache_prompt and generation cache options; capture cache state during benchmarks.
- **yes · documented**: Declares a configurable HTTP server interface. [MLX-LM server interface](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/mlx_lm/server.py) Scope: Repository-declared host and port interface; not executed.
- **yes · documented**: Declares a /v1/chat/completions route. [MLX-LM server interface](https://github.com/ml-explore/mlx-lm/blob/872ae88d1fac77350db23c8c04fe8dd372a9e3e8/mlx_lm/server.py) Scope: Repository-declared route only; protocol compatibility and client behavior untested.

## Open questions

- Which pinned model, quantization, context length and runtime versions fit the selected Mac?
- What are measured latency, throughput, peak memory and swap under one and four concurrent requests?
- Concurrency and JSON schema output are Unknown in this research package; no cited absence is asserted.

## Verification plan

1. Choose hardware and obtain explicit approval for installations, downloads and execution resources.
2. Pin runtime and engine versions, model artifact hashes, tokenizer, template, quantization and inference settings.
3. Run the local-model-mac investigation with synthetic prompts and capture raw measurements; do not record documentation as execution.

Last checked: 2026-09-15. Recheck by: 2026-10-15.

Content rights: Unlicensed prototype; redistribution decision pending.