Docs
Models & downloads
Discover, import, download, and manage models from Hugging Face — all local.
Bedrova’s library is where you find, download, and manage local Mac-runnable models.
Discover
Browse curated starter and power-user picks, or search Hugging Face directly by keyword,
user/model, or full URL. Each result shows its model card (README), download variants
(MLX / GGUF / provider pack), size, and a compatibility read for your Mac. When an MLX and
GGUF text variant are both viable, MLX is recommended and GGUF is labeled broad
compatibility, usually slower. Gated or remote-code repos are flagged before you download.
Screenshot: the Discover tab with search + curated picks — added at launch.
Download
Downloads are resumable with per-shard progress, free-space preflight, and SHA-256 verification. Pause, resume, cancel, or repair a partial download at any time. Add your Hugging Face token (stored in the Keychain) for gated models and faster limits.
Import local models
Already have an MLX model folder, GGUF text model, or Core ML perception package? Import it
directly — Bedrova inspects it, reuses any bedrova.model.json sidecar where present, and
adds it to your Local Models with trust state and a repair action if anything’s missing.
Manage
Every model has a profile: system prompt, load settings, sampling, reasoning, structured output, speculative decoding, and prompt template — layered over your global defaults. Pin favorites, set per-model defaults, or unload to free memory.
Supported families
Text: DeepSeek (incl. V4 Flash), Qwen, Gemma, Llama, Mistral/Mixtral, MiniMax M2, and more. GGUF text models can run through the secondary llama.cpp pack when selected. Vision: Qwen-VL and other MLX-VLM models. Embeddings/search: Nomic, BGE, Qwen3, Gemma, MiniLM and provider-specific multimodal search models where supported. Perception: Core ML packages for classification, detection, segmentation, depth, OCR, and pose. Voice and image models are managed the same way (see Voice and Images). Some newer architectures stay capability-gated until a native Mac runtime exists — Bedrova tells you when a model can’t run yet.