MLX-VLM 0.7.9: FLUX 3, MOSS and Nemotron on Mac

MLX-VLM 0.7.9 packages support for FLUX 3 image generation and editing, MOSS-Image, and NVIDIA's Nemotron 3.5 Super VL and Bolt. For developers running these models on a Mac, the first checks are the supported checkpoint format, available storage and memory required by the intended workload.

The stable package release was published on October 19. The model integrations merged on October 16, so this is their packaged runtime availability. The models themselves were announced separately. The release also adds image-resolution controls for the server and GRPO training with optional vision inputs.

The Python package requires Python 3.10 or newer, with dependencies including MLX 0.32.4 or newer and Transformers 5.14 or newer. Start in a separate environment if an existing project pins older versions, and keep a record of the resolved packages alongside the model checkpoint.

Choose the FLUX 3 checkpoint before downloading

The versioned FLUX 3 guide supports standard BF16 checkpoints. Its CUDA FP8 and NVFP4 variants are unsupported by this implementation. A smaller download in another format therefore needs its own compatibility check.

FLUX 3 Image Klein can generate an image or edit using up to ten reference images. Prompt expansion runs automatically for both tasks, and the expanded prompt is returned with the result. Save that prompt when comparing outputs, since it records what the image model received after expansion.

The standard Klein model defaults to 16 steps with guidance 1. The base version defaults to 50 steps with guidance 3.5. Keep those settings associated with the correct checkpoint when moving a recipe between machines.

The image model and prompt upsampler together require approximately 39 GB of disk space, according to the guide. Hugging Face access conditions and account authentication must be handled before the first download. Requested dimensions must be multiples of 16, from 64 to 4096 pixels per side; an edit without an explicit size follows the first reference image.

Set MOSS image dimensions explicitly

MLX-VLM supports both MOSS-Image-Base and MOSS-Image-SFT using their original Hugging Face checkpoints. The MOSS integration guide describes inference without PyTorch or Diffusers at runtime.

Both checkpoints default to 50 sampling steps, guidance 3.0 and scheduler shift 5.0. The generic image API defaults to 512 by 512 pixels, while the upstream recommendation is 1024 by 1024. Set the dimensions deliberately when reproducing a sample or comparing output quality. Both dimensions must be positive multiples of 16.

The original MOSS weights occupy approximately 27 GB on disk. The runtime releases the text encoder before loading the diffusion transformer by default. That helps explain its loading sequence, but the download size alone cannot establish how much memory a job will need.

For either image integration, begin with one image at the intended resolution and record peak memory and generation time. Increase the batch or reference-image count only after that baseline works. The 39 GB and 27 GB figures above describe storage, so they should not be used as RAM requirements when choosing a machine.

Nemotron support has different loading paths

The Nemotron runtime guide documents direct loading of the Super VL NVFP4 checkpoint. It preserves quantized expert weights while decoding FP8 dense layers to BF16. The checkpoint occupies about 75 GB on disk, with additional memory needed for loading, vision processing and the context cache.

The guide describes short-context Super VL requests on a 128 GB Mac. Treat that as a documented configuration with a limited workload, rather than a guarantee for longer conversations or larger visual inputs.

Bolt has a different route. Its BF16 checkpoint is approximately 150 GB before inference caches, and the documented MLX four-bit conversion occupies about 43.1 GB on disk. Use the size and format of that specific conversion when planning a local deployment; figures for other Bolt variants do not describe this artifact.

Super VL's bundled multi-token prediction head supports speculative decoding. The released Bolt checkpoint has no MTP tensors and cannot act as its own native drafter. For Super VL video MTP, the guide requires video_pruning_rate=0; pruned video supports ordinary generation. Image and video inputs also need separate requests, and Super VL and Bolt have no audio encoder.

Recheck server and training recipes

The image-resolution change lets server defaults for min_pixels and max_pixels change through PATCH /v1/settings without reloading the model. Updates apply to subsequent requests. A per-request value takes precedence, while already queued generation retains its resolved settings.

For an OCR service, this gives a practical way to test a higher-detail input setting while keeping the model loaded. Compare the result and token use on representative documents before changing the default for all requests. A resolution control does not establish a quality improvement on every document or model.

The training guide adds GRPO recipes for text and image-conditioned responses. Training needs the optional training dependencies. The shared Trainer takes an explicit optimizer, with learning rate and weight decay configured on that optimizer. Audio and video GRPO are not currently supported.

MLX-VLM's framework is MIT-licensed, while model permissions remain separate. FLUX 3 image weights use the FLUX Kommunity License; the MOSS guide identifies Apache 2.0 terms for its weights and upstream implementation. Check the chosen checkpoint's terms and access requirements before distributing it or building a commercial workflow around it. Local hardware and storage costs remain part of that deployment.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool and AI infrastructure teams explain releases, deployment requirements and technical tradeoffs through useful content and distribution.

Book a 30-minute call.