Models
Catalogue
Section titled “Catalogue”- Built at release by
scripts/build_catalog.pyfrom live Hugging Face metadata and bundled as a snapshot. At runtime minibrain never calls Hugging Face for the catalogue; the optional sync reads only our own catalogue URL (MINIBRAIN_CATALOG_URL). - Rules: approved publishers (Qwen, Google, Meta, Mistral, Microsoft, IBM, NVIDIA, DeepSeek, OpenAI gpt-oss, Allen AI, HuggingFaceTB, Z.ai GLM), instruct or chat variants, reviewed commercial licences, architectures our engines support, safetensors present.
- GLM: GLM-4-0414 (
glm4) and GLM-4.5 / 4.6 (glm4_moe) run on transformers 5, SGLang, mlx-lm and llama.cpp. GLM-4.5 and 4.6 are hybrid reasoning models; minibrain turns thinking off through the chat template. GLM-4.5-Air and 4.6 are XL (mixture of experts, 106B and 355B). - Every generation stays; older ones show “Newer version available”.
- Size tiers: S ≤4B, M 7–14B, L 24–35B or a small mixture of experts, XL 70B+ (multi-GPU).
- Default order: recommended for the selected pool; also by popularity, size or name. Filters: hardware fit levels on
the selected pool, latest generation only, download access, publisher, task, size, format (full precision or
quantized) and licence (long groups show five options and “Show more”), with removable chips. Columns can be shown
or hidden; a row opens the model’s page (
/discover/<id>): formats with fit, requirements, licence and files. - Formats read “Full precision · BF16” or “Quantized · Q8_0” / “Quantized · FP8”. Quantized formats are serve only.
- Catalogue updates: Settings → General → Model catalogue (Check for updates, or every day).
Add from Hugging Face
Section titled “Add from Hugging Face”Paste a link or id; main reads that one repository’s metadata, on request, from the configured endpoint (Hugging Face
or a mirror), reads config.json and adds the model to Discover. Only when an admin sets the Model
sources policy to Any; the default, Approved publishers, keeps everyone on the curated list.
Sources
Section titled “Sources”| Source | Configured by |
|---|---|
| Hugging Face or a mirror | MINIBRAIN_HF_ENDPOINT (or HF_ENDPOINT) |
| S3-compatible bucket | A full s3:// URL, or a path under MINIBRAIN_MODELS_S3_URL; AWS_ENDPOINT_URL_S3 for non-AWS stores |
| Folder on the server | A path |
Weights are pinned by repository and revision and fetched by a download or import job.
Import formats
Section titled “Import formats”- A Hugging Face format folder:
config.json, the tokenizer and.safetensorsweights, sharded or not. - Pickle weights (
.bin,.pt,.pth,.pkl,.ckpt) are refused: they can run code when loaded. - An import whose revision and files match a catalogue variant is verified and takes the catalogue’s details; anything else is listed as unverified.
- GGUF files run on llama.cpp, base models only (engines).
- Quantized weights (GGUF, or safetensors whose
config.jsonhas aquantization_config, such as FP8) serve but are not fine-tuned: transformers refuses to train FP8 layers on the GPUs that run them natively. Training a quantized base is refused up front; train on its full-precision format.