Image
Make images
FLUX, Stable Diffusion and Qwen Image, generated on your Mac.
Native macOS · Apple Silicon · runs side by side
Image, video, speech, vision, text. Every kind of AI model, on your Mac.
Run FLUX, Wan, Whisper, DeepSeek, Qwen and the latest open models on-device, side by side, through one API. MLX-first on Apple Silicon, with Core ML perception and GGUF text compatibility.
- by ideius
- no telemetry
- notarized
- Apple Silicon
- free
One app, every type
Stop running an app per model type.
A chat app, an image tool, a transcription app, a video tool. Bedrova runs all of them on-device, from one library.
Video
Make video
Text-to-video and image-to-video, rendered on-device. Nobody else does this locally.
Chat
Chat and reasoning
Llama, Qwen, Gemma and DeepSeek V4 Flash, running locally.
Voice
Speech in and out
Whisper transcription and natural speech, no cloud.
Vision
See images
Ask about screenshots and photos.
Perception
Read and measure
OCR, detection, segmentation, depth and pose through Core ML.
Embeddings
Chat with your documents
Local embeddings, retrieval and citations you can click.
Local API
OpenAI-compatible server
Point any app or agent at localhost:1337/v1.
The Loom
One conversation. Several models at once.
Most local AI apps give you one model and one thread. The Loom gives you a conversation that spans text, vision, voice and image models together. Paste a screenshot and ask about it, have the answer read aloud, generate an image from it, and keep going in the same thread.
- Each model type runs in its own lane, so you can see what's working.
- Lanes that can run together do. Lanes that can't say so, plainly.
- Tool calling and guaranteed-shape JSON, on any lane.
One honest caveat: the largest reasoning models take the whole GPU, so they run a turn at a time while other lanes carry on. Bedrova shows you that rather than animating a concurrency it doesn't have.
The runtime
One API. Every type. Side by side.
Your apps make one kind of call. Bedrova sends each request to the right model. Favourites stay pinned and always loaded, everything else loads on demand, and your Mac's memory is budgeted across every loaded model.
Your apps & agents
Chat Qwen3.5 7B always-on
Image FLUX.1 schnell always-on
Voice Whisper v3 always-on
Vision Qwen2.5-VL JIT
Embeddings Nomic Embed JIT
RAM budgeted across loaded models — never OOMs your Mac
Qwen3.5 · 4.3 GBFLUX.1 · 11 GBWhisper · 3.1 GB 13.6 GB free
Getting started
Running in about three minutes.
No Python, no terminal, no config files. An 8 GB Mac is enough to start.
- 1
Open it
Bedrova reads your chip, memory and free disk, then recommends a model that actually fits.
- 2
Fetch a model
One click downloads and verifies it. Resumable, and safe to close mid-download.
- 3
Start talking
Type. Add more model types later, whenever you want them.
Built for your Mac
Built for Apple Silicon, not ported to it.
- Tuned for Apple Silicon with MLX, the fast path on M-series chips.
- Runs frontier models like DeepSeek V4 Flash entirely on your Mac.
- Real control over quants, context, sampling, mixture-of-experts and speculative decoding.
- GGUF models work too. Broad compatibility, usually slower, and labelled that way.
- We name the engines we build on: mlx-lm, mflux, mlx-audio, llama.cpp, Core ML.
Control + safety
Run big models without nuking your Mac.
Bedrova tracks memory across every loaded model and checks a new one against what's left before loading it. Either it fits, or you get told why not and what to try instead.
Qwen3.5 · 4.3 GBFLUX.1 · 11 GBWhisper · 3.1 GB 13.6 GB free
FLUX.1 dev · 24 GB — won't fit → try FLUX.1 schnell · 11 GB
Your Mac is
the whole datacentre.
- Private — Prompts, files and models never leave your Mac.
- Offline — Works on a plane, in a basement, anywhere.
- Free and unlimited — No per-token bills, no rate limits.
- Yours — No account. Nobody's permission needed.
Every kind of AI model, running on your Mac.
Free, native, and entirely on-device. One runtime for all of it.