Model source

HuggingFace models

Studio can search the HuggingFace Hub for GGUF repositories, inspect what is inside, download with resume, and convert the result to a packaged .smm. Public downloads need no account.

GGUF browser licence + gated status resume convert to .smm

What the browser does

Search the Hub for GGUF repositories instead of guessing URLs.
Show licence, gated status, files, quantisation and size before downloading.
Download to %USERPROFILE%\SharpMind\Models with progress and resume.
Convert a downloaded GGUF to .smm from the same panel.

The browser is under Training → Models. It does not install a separate package or require a HuggingFace account for public files.

Token

A token is only needed for gated repositories. Add it under Settings → Providers → HuggingFace Model source. It is stored encrypted with the same local key box as the Compute and OpenRouter secrets.

Public GGUF downloads work with no token and no username.
A gated repository also requires accepting its terms on the Hub with the same account.
Delete the token from the same panel; it is not written to model files.

Browse and download

Training → Models → Find GGUF models.
Search by repository or model name. Filter by what you need: size, quantisation, licence.
Open a result to see the file list. Pick the quantisation that fits this machine or host.
Download. If it stops, press download again; the partial file resumes.
Convert to .smm when you want the full SharpMind path, a Compute host upload or a fine-tune base.

Convert to .smm

The converter sits directly under the browser. It suggests the destination, verifies the packaged result and refuses to overwrite without confirmation.

source:      model-Q4_K_M.gguf
destination: model-Q4_K_M.smm
verify:      read back config, weights and tokenizer metadata

A Host upload and a fine-tune base both need .smm. For a local chat that already loads the GGUF, conversion is optional.

Pick a quantisation

Quantisation	Trade-off	Use when
Q8	Largest, closest to full precision	You have memory to spare and want quality.
Q5 / Q6	Good balance	The model fits but Q8 is tight.
Q4	Smallest useful files	Laptop inference, quick evaluation.
Q2 / Q3	Very small, visibly degraded	Only when nothing else fits.

Context length also costs memory. A model that "fits" at a short context may fail at a longer one; the chat badge is advisory.

Use the model

Local chat: pick the model in Chat; no upload.
Compute host: host card → Browse → Upload and use.
Fine-tune: the base must be .smm, so convert first.
Work: use a tool-capable local model; small quantised models are weaker at tool calls.

Problems

Message / symptom	Fix
401 or "gated repository"	Add a token and accept the repository terms on the Hub with that account.
403 forbidden	The licence is not accepted or the account lacks access. Check the repo page on the Hub.
Download restarts from zero	The server did not return a range, or the partial file was removed. Resume needs the same destination path.
No space left on device	Models are large; free disk in Models or download to another drive.
Conversion fails	The file may be corrupt or an unsupported GGUF variant. Re-download or try another quantisation.
Model downloads but quality is poor	Try a higher quantisation; Q2/Q3 can be visibly degraded.

Models & chat → Fine-tune a base Troubleshooting