testedknowhowLAB

Find a space, target or capability

Search every research space and the entities inside them.

Inference library / PRODUCT DOSSIER

MLX-LM

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

Investigation: Which local model runtime fits a 32GB or 64GB Mac? →

Official website ↗Explore in graph →8 cited claims ↓
Local Apple Silicon execution with MLX; model conversion and server configuration remain operator-owned. · Open source · MIT runtime source; model weights have separate licensesRevision 2026-09-15.1 · unreviewedNeeds recheck · recheck 2026-10-150 attestations · checked 2026-09-15

What we do not know yet · 3 open gaps

  • 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.

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

Applicability, 2 recheck triggers & verification plan

Applicability

Repository revision 872ae88d1fac77350db23c8c04fe8dd372a9e3e8; living documentation separately dated.

Local Apple Silicon execution with MLX; model conversion and server configuration remain operator-owned.

Excludes: Runtime verification and performance rankings; Model weights license or quality evaluation; Cloud inference, remote hosts and external tools; Unlisted model architectures, builds and hardware

Recheck when

  • Runtime, bundled engine, backend, model format or license changes.
  • Documentation changes affecting the cited configuration.

Needs recheck · due

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.

Provenance & review

Testedknowhow model runtime research

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

Revision 2026-09-15.1 · unreviewed · checked

0 attestations · Content: Unlicensed prototype; redistribution decision pending

Capability claims · 8 / 14 sourced

Standing condition: these are documentation claims at the source capture date, with outcome “claimed”; no runtime execution is recorded. Dossier applicability and limitations apply throughout. Expand a row for its statement, specific scope and contradictions. Numbered citations resolve to the source locator and capture date below. Missing evidence means Unknown, not No. Sources are unpinned unless marked “Pinned”; this marks source stability, not stronger support or runtime verification. Reuse counts refer to claims in the full dossier; repeated citations do not represent independent sources.

Yes Documented within scope Partial Limited form of the feature No Explicit evidence of absence Unknown No evidence recorded. These are support assessments, not quality scores or runtime verification.

Execution & hardware

Local inferenceYes
Explore Local inference →

Provides local text generation with MLX.

Scope: Model files must be acquired before an offline run.

[1]
Apple Silicon accelerationYes
Explore Apple Silicon acceleration →

Targets Apple Silicon through the MLX framework.

Scope: Supported Apple hardware and MLX installation; speed unmeasured.

[1]

Model formats

MLX modelsYes
Explore MLX models →

Loads MLX-compatible model repositories or local conversions.

Scope: Supported architectures and MLX-compatible weight layouts.

[1]
Quantized weightsYes
Explore Quantized weights →

Provides conversion to quantized model weights.

Scope: README conversion example uses 4-bit weights; not equivalent to GGUF 4-bit artifacts.

[1]

Serving

Local HTTP serverYes
Explore Local HTTP server →

Declares a configurable HTTP server interface.

Scope: Repository-declared host and port interface; not executed.

[2]
OpenAI-style chat endpointYes
Explore OpenAI-style chat endpoint →

Declares a /v1/chat/completions route.

Scope: Repository-declared route only; protocol compatibility and client behavior untested.

[2]
Streaming generationYes
Explore Streaming generation →

Provides incremental output through stream_generate.

Scope: Python streaming interface; no blanket HTTP API parity claim.

[1]

Runtime controls

Prompt cache reuseYes
Explore Prompt cache reuse →

Provides prompt caching for repeated long prompts.

Scope: mlx_lm.cache_prompt and generation cache options; capture cache state during benchmarks.

[1]

Unknown · 6 capabilities without evidence

Interface images & local attachments
Screenshots & attachments
Image index

Official project screenshots and your own reference images. Images illustrate a UI; they do not verify feature behavior. Uploads stay in this browser.

No screenshots gathered yet.

Sources · 3

[1]MLX-LM documentation ↗Pinned6 / 8 claims
repository · locator & rights

872ae88d1fac77350db23c8c04fe8dd372a9e3e8

Quick Start; Python API; quantization; long prompts and generations

Link and original paraphrase only; upstream rights retained.

[2]MLX-LM server interface ↗Pinned2 / 8 claims
repository · locator & rights

872ae88d1fac77350db23c8c04fe8dd372a9e3e8

main argument parser: MLX Http Server, --host, --port; /v1/chat/completions route

Link and original paraphrase only; upstream rights retained.

[3]MLX-LM license ↗Pinned0 / 8 claims
repository · locator & rights

872ae88d1fac77350db23c8c04fe8dd372a9e3e8

MIT License

Link and original paraphrase only; upstream rights retained.