MLX-VLM
Blaizzy (Prince Canuma)MLX-VLM runs vision-language and omni models locally on Apple silicon using MLX, taking image, audio and video inputs alongside text. It ships a CLI and a FastAPI server with continuous batching, automatic prefix caching, KV-cache quantization and speculative decoding, and converts and quantizes Hugging Face models for local use. Fine-tuning with LoRA and QLoRA is included but the package is inference-first.
Reached Round 1 as a finetuning_code candidate and was refiled here per ledger #431; the README is inference-first by a wide margin.
Openness
5 high confidence- license
- MIT(OSI)
- source
- public
- core features withheld
- no — no enterprise or paid tier in the repository or README
MLX-VLM's LICENSE is the stock MIT text with no appended condition, and the public repository is the package itself. Nothing in the README or repository offers a hosted, paid or enterprise build, so the core is ungated.
- https://api.github.com/repos/Blaizzy/mlx-vlm recorded 2026-09-01
Repo metadata for Blaizzy/mlx-vlm - archived false, license MIT, 5,454 stars, pushed 2026-09-01.
- https://raw.githubusercontent.com/Blaizzy/mlx-vlm/main/LICENSE recorded 2026-10-07
LICENSE body, read in full: stock MIT text, "Copyright (c) 2025 Prince Canuma", no appended condition.
- https://raw.githubusercontent.com/Blaizzy/mlx-vlm/main/README.md recorded 2026-10-07
README, read for a paid tier or gated build (none found) and for the product's shape: "MLX-VLM is a package for inference and fine-tuning of Vision Language Models (VLMs) and Omni Models ... on your Mac using MLX"; FastAPI server, continuous batching, automatic prefix caching, KV cache quantization, speculative decoding, distributed inference; fine-tuning is one late section.
Adoption
3 high confidenceMLX-VLM's adoption reading comes from PyPI downloads of the mlx-vlm package over the trailing 30 days; the package's metadata links back to the Blaizzy/mlx-vlm repository, confirming it is the right artifact to measure. As a download count, it measures package fetches rather than distinct users or installs.
- https://pypi.org/pypi/mlx-vlm/json recorded 2026-09-01
project_urls Homepage/Repository = https://github.com/Blaizzy/mlx-vlm; version 0.6.17; summary names inference and fine-tuning of VLMs on Mac using MLX.
- https://pypistats.org/api/packages/mlx-vlm/recent recorded 2026-09-01
"last_month":777221
Capability
3 medium confidenceMLX-VLM sits two tiers below vLLM, level with uzu and litert-lm among the other single-hardware local engines: it targets one hardware family and lacks the multi-node, multi-accelerator parallelism of the engines above it. Its multimodal input coverage, image, audio and video, is the broadest among these local engines.
- https://raw.githubusercontent.com/Blaizzy/mlx-vlm/main/README.md recorded 2026-10-07
README sections: Server (FastAPI), Continuous Batching, Automatic Prefix Caching (APC), KV Cache Quantization, Speculative Decoding, Distributed Inference, Fine-tuning.
Verified 2026-09-24