brain

brain trains and runs neural networks on the hardware you already have.

One Rust program does the whole job: download a model, run it, fine-tune it, quantize it so it fits, and serve it behind an OpenAI-compatible API. It needs no Python, no PyTorch and no separate inference runtime, and the same models run on NVIDIA cards, on other GPUs, on a plain CPU, on an Intel NPU or in a browser.

Built by Swedish Embedded AB - we put AI on hardware that ships. Hire us.

Three projects, one stack

brain is one of three open projects that work together. Each one stands on its own:

Project	Its role
brain (this repository)	Runs the models. Inference, training, fine-tuning, quantization and serving.
sven	Runs the agent. A terminal AI agent that reads code, runs tools and works through tasks, using any model provider, brain included.
Splinter	Turns agent experience into better local models. It runs sven on real tasks, checks the results, trains a model with brain, and releases it only when held-out tests show it improved.

You can use brain without the other two: it serves models to any application that speaks the OpenAI, Anthropic or OpenRouter API, or D-Bus, or links the Rust SDK.

What brain does

	Models	What you can do
Text	Qwen3, Qwen3.5, GLM, GPT, LFM2.5	chat, tool calls, embeddings; serve many users at once; train from scratch; fine-tune with LoRA
Vision	YOLOv8, SAM 2.1, ZipDepth, SCRFD, ArcFace, CLIP, Florence-2	object detection, segmentation and video tracking, depth, face detection and recognition
Vision and language	Qwen3-VL, MiniCPM-V, FastVLM, Moondream 3, DeepSeek-OCR, Qwen3-Omni	describe images, answer questions about them, turn documents into markdown
Image and video generation	Z-Image, FLUX.2 Klein, FLUX.1 Kontext, SDXL, Wan, LTX	text to image, editing, inpainting, identity control, LoRA training, text to video
Speech and audio	Qwen3-TTS, CosyVoice, Nemotron ASR, Qwen3-ASR, MiniMax Music	text to speech with voice cloning, speech to text, music
Forecasting, 3D, world models	Chronos-2, Kronos, WorldMirror-2, DIAMOND	time-series forecasts, photos to 3D scenes, playable simulations

Every model is used the same way: brain <model> <action> on the command line, and the same models over HTTP, D-Bus and the SDK. The catalog, with a page per model saying what it supports and what it does not yet, is docs/reference/models/index.md.

Hardware

brain picks up whatever compute the machine has; brain devices shows what it found. Four backends are compiled into every native build and share the same kernels; pick a device with --device and a backend with --backend:

Backend	Runs on
wgpu (default)	any GPU with a Vulkan, Metal or DirectX 12 driver: NVIDIA, AMD, Intel, Apple; also WebGPU in a browser
vulkan	Vulkan driven directly
cuda	NVIDIA GPUs through the driver, with native CUDA kernels for the hottest paths (details)
cpu	the CPU's cores; slower, but nothing else needed

The Intel NPU is a separate path for a subset of models: they are exported to ONNX and run through OpenVINO rather than on brain's own kernels, and it is opt-in per model (each model page says whether it has one). A model too big for one card can be split across several, quantized, or streamed from memory as it runs (docs/guides/hardware/choosing-hardware.md).

Get started

Build it (a recent stable Rust toolchain, plus CMake and a C compiler for the bundled libopus):

make build/release
export PATH="$PWD/target/release:$PATH"     # once per shell
brain doctor                                # checks the model store, the disk and the GPUs, and says how to fix what fails

Run a model. Its weights download the first time it is used:

$ brain infer qwen3 --prompt "The capital of France is" --max-new 12
The capital of France is Paris. The capital of Italy is Rome. The capital of

Serve it behind an OpenAI-compatible API (it also speaks the Anthropic and OpenRouter APIs and D-Bus):

$ brain serve --openai 8799 &
$ curl http://127.0.0.1:8799/v1/chat/completions -H "Authorization: Bearer $APIKEY" \
    -H 'Content-Type: application/json' \
    -d '{"model":"Qwen/Qwen3-0.6B","messages":[{"role":"user","content":"Say hello in exactly five words."}]}'

docs/getting-started/quickstart.md trains a small model from scratch and serves a real one in about five minutes, and docs/getting-started/tour.md shows real commands and their output, one model feeding the next.

How you know it works

brain does not lean on another framework to be correct, so it checks itself:

• Every imported model is compared with its reference implementation, step by step, by tests you can run.
• Training is checked numerically against finite differences (make test/gradcheck).
• Backends are checked against each other (make test/parity).
• Limits are written down. Where a model is not yet trainable, not yet servable or not checked at full size, its page says so.

Documentation

Everything is under docs/ and is published at https://swedishembedded.com/docs/brain.

Start	docs/index.md, install, quickstart
Guides	inference, training, hardware, serving
Concepts	architecture, scaling across GPUs
Reference	CLI, SDK, configuration, models
Contributing	docs/contributing/index.md, AGENTS.md

Who builds brain

brain is built by Swedish Embedded AB. We build AI that runs on hardware that ships: the GPU you already have, a CPU with no GPU at all, an Intel NPU, a board in the field or a browser tab. Everything in this repository is that work, done in the open and held to tests you can run yourself.

If your team needs any of this, you can hire us:

• Running models on the hardware you have - NVIDIA and other GPUs, CPUs, Intel NPUs, embedded Linux boards, the browser.
• Making a large model fit - quantization, weight streaming, splitting a model across GPUs.
• Porting a model from a paper or a PyTorch checkpoint, with its results checked against the original.
• Making GPU code fast - and proving it is still correct afterwards.
• Production inference - serving many users at once, keeping models loaded, scheduling across accelerators.
• Embedded and real-time firmware alongside the AI, which is where this company started.

Send an email to info@swedishembedded.com and tell us what you are trying to ship.

License

Apache-2.0 - see LICENSE. Model weights carry their own licenses; see docs/compliance/third-party-models.md.

Copyright (c) 2026 Martin Schröder.