MEGA Hub
New
Docs/Models

Model Repositories#

Model repositories keep weights, configuration, tokenizer assets, documentation, and release history under one versioned owner/name identity. They use the same commit and access model as every MEGA repository, with model-specific discovery and download entry points.

Create a model repository#

Use New Model in the browser or create one with the CLI:

BASH
mega repos create alice/qwen-demo \  --type model \  --description "Qwen checkpoint packaged for MEGA" \  --license apache-2.0 \  --tag text-generation \  --tag 0.8b

Repositories are public unless --private is passed or an organization policy chooses a private default. The repository type cannot be changed after creation.

A MEGA-native model release normally includes:

qwen-demo/├── README.md├── config.json├── tokenizer.json├── model.mega.index.json├── model-00001-of-00002.mega└── model-00002-of-00002.mega

README.md is the model card shown on the repository page. The index is the stable runtime entry point for a sharded model. Keep configuration and tokenizer files beside the artifact so a pinned revision is self-contained.

See Repository Cards for documentation guidance and Artifact Format for the binary and index contracts.

Metadata and discovery#

The root README.md may declare Hugging Face YAML metadata. Card description and license values override repository-setting fallbacks, while card and repository tags are merged. Visibility remains an access-controlled repository setting. See Repository Cards for the complete precedence and validation rules.

Publish reproducible benchmark evidence with Model Evaluations. Keep the score, metric, task, source, and release revision aligned with the model card.

On the model page, repository relationships are grouped as a model tree, Spaces using the model, Collections including the model, and paper or article references. The inference panel is separate: it reflects live model-to-Provider routes rather than card tags. MEGA-native artifacts also receive a MegaTensors metadata summary from bounded inspection of the selected .mega file or .mega.index.json.

When a healthy chat-completions route exists, the same panel includes a signed-in Model Widget. It uses the account's MEGA Inference balance, supports Provider selection, and keeps bounded multi-session history in account-scoped local browser storage. Model publishers cannot enable the widget by adding card markup; the live Provider mapping is authoritative.

Fallback metadata can be updated without rewriting the card:

BASH
mega repos settings alice/qwen-demo \  --description "Instruction-tuned 0.8B checkpoint" \  --license apache-2.0 \  --tag text-generation \  --tag qwen \  --tag 0.8b

Model search matches the repository ID, effective description, tags, and license. Task and library signals can also come from parsed card metadata. Use specific, stable tags rather than repeating prose from the card.

An inference task tag does not make a model callable by itself. Live Router availability is controlled by the separate model-to-Provider mapping shown at Inference Models. See Inference Providers for routing and billing.

Upload a release#

Upload a directory through the typed model command:

BASH
mega models upload alice/qwen-demo ./release . \  --commit-message "Publish MEGA release"

For a large release tree, the resumable uploader preserves completed multipart parts between attempts:

BASH
mega upload-large-folder alice/qwen-demo ./release --revision main

You can also push over Git. Large-object transfer is handled automatically; never commit credentials, local caches, or private signing keys.

Download and load#

Download a complete revision or selected files:

BASH
mega models download alice/qwen-demo --revision main --local-dir ./modelmega download alice/qwen-demo config.json tokenizer.json --local-dir ./model

For reproducible automation, replace main with a release tag or commit ID. Python can download from the Hub and then open the pinned artifact:

PYTHON
from megatensors.hub import MegaHubClientfrom megatensors import mega_open client = MegaHubClient()root = client.snapshot_download("alice/qwen-demo", revision="v1.0") with mega_open(root / "model.mega.index.json", device="cpu") as artifact:    print(len(list(artifact.keys())))

Use the Python SDK for the current download method signatures.

Release checklist#

  • Add a model card with intended use, limitations, inputs, outputs, and provenance.
  • Declare a license and task-oriented tags in repository metadata.
  • Keep the model index, every referenced shard, configuration, and tokenizer in the same revision.
  • Verify loading from a clean directory before creating a release tag.
  • Pin deployments to a tag or commit instead of main.
  • Use Signing and Trust when consumers require publisher provenance.
  • Use Gated Repositories when access requires an application or license acknowledgement.