Frequently asked questions

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique. Instead of updating all model weights, it injects small trainable rank-decomposition matrices into each transformer layer. The adapter file stores only those matrices — typically 1–200 MiB compared to gigabytes for the full model. At inference time the adapter weights are merged with the frozen base model.

safetensors is a fast, safe serialisation format for PyTorch/JAX tensors. It is used with PEFT/diffusers and loaded with PeftModel.from_pretrained(). GGUF (GPT-Generated Unified Format) is used by llama.cpp and Ollama for CPU/GPU inference with quantisation baked in. Most text-generation adapters in this registry are published in both formats. Image adapters (SDXL, FLUX) are safetensors-only.

It depends on two licences: the adapter's own licence and the base model's licence. For example, adapters on meta-llama/Meta-Llama-3-8B-Instruct are governed by the Llama Community Licence (commercial use permitted for most companies; >700M MAU requires Meta approval). Always check the License field on the adapter detail page and the base model card on Hugging Face.

Rank (r) controls the size of the injected matrices: a rank-8 adapter is smaller and trains faster; rank-64 captures more nuance but adds more parameters. Most adapters here use r=8 or r=16 for style transfer and r=32–64 for instruction following. The rank is fixed at training time and cannot be changed without retraining.

Download the GGUF adapter, then create a Modelfile:

FROM llama3
ADAPTER ./my-adapter.gguf

Then run ollama create my-model -f Modelfile. The mfl pull CLI generates this file automatically when you pull a GGUF adapter.

Create an account, then use the CLI:

mfl login
mfl push ./my-lora.safetensors \
  --name "My adapter" \
  --base-model stabilityai/stable-diffusion-xl-base-1.0 \
  --task text-to-image \
  --license apache-2.0

Adapters are private by default. Toggle visibility under adapter settings once uploaded.

Public adapters: 2 GiB per file. Members with a paid plan: 10 GiB. If your GGUF quantisation produces a file larger than 2 GiB, split it into shards with llama.cpp/gguf-split and upload each shard separately.

Yes. Every artifact is scanned with ClamAV and an internal pickle-safety checker on ingest. safetensors files are always safe by design (no executable code). GGUF files are validated against the spec. We do not run uploaded model weights.

Generate a token on the API tokens page, then run:

mfl login --token mfl_your_token_here

The token is stored in ~/.config/mfl/auth.json. You can also set MFL_TOKEN as an environment variable for CI pipelines.

Unauthenticated requests: 60 per minute per IP. Authenticated requests: 600 per minute. Download endpoints count separately and are not rate-limited, but concurrent downloads per account are capped at 4. Exceeding any limit returns HTTP 429 with a Retry-After header.