Why we started Synthetic Minds
Synthetic Minds is an open-source AI research and engineering lab. Our mission is to make Neural Accelerators an open, accessible foundation for AI, so that anyone can train, fine-tune and run models without being locked into a single vendor’s stack. Today we’re sharing why we started the company, and what we believe.
The cost of the old way
For years, GPU compute has been the default for AI. It still works, but it is showing its age. In our experience, GPU prices keep rising year over year, and compute is more expensive than ever. As demand outpaces supply, access has become harder for the developers who need it most: independent researchers, startups and small teams who cannot simply write a very large cheque.
When the barrier to building with AI is the price of the hardware, fewer people get to build. We think that is a problem worth solving.
The next step is hardware designed for AI
We believe the next step in artificial intelligence is the move from general-purpose GPUs to Neural Accelerators: hardware designed for AI from the start. That is more than a change of chip. It means rethinking how models are designed, adapted and run, to take advantage of how these machines actually work instead of translating workflows made for something else.
That rethinking needs new software. The tools for training, fine-tuning and running open models on Neural Accelerators are still early, and that is the gap we are here to close.
Why Tenstorrent
We chose Tenstorrent, one of the first Neural Accelerator makers to sell to both the public and enterprises, for three reasons.
- An open software stack. We can read it, build on it and contribute back.
- Open kernels. We can work close to the hardware instead of treating it as a black box.
- Freedom from vendor lock-in. We are not tied to CUDA or ROCm, so we can design the platform we think is right.
That freedom is what lets us aim high: a tightly integrated platform for training and running AI models, built from the ground up on Tenstorrent Blackhole and Galaxy. Both power our own infrastructure and frameworks. We are building toward one of the highest-performing platforms we can, and we will publish performance numbers only once they have been measured on real silicon.
What we are building
Our first project is Infall-TT, an open-source framework for training, fine-tuning and running open models on Tenstorrent Neural Accelerators. It is in development and will bring quantization to Tenstorrent’s native block-float formats, LoRA fine-tuning on a frozen 4-bit base, and export to the formats people already use. You can join the waitlist to be reminded when the beta opens.
How we work
- Open by default. We release the core of everything we build under the Apache 2.0 license, on GitHub, with models and adapters on Hugging Face as each one is ready.
- Native, not ported. We design our tools around the hardware instead of translating GPU workflows.
- Correctness first. Every operation is tested against a reference implementation with explicit tolerances.
Careers
We are a small team, and we would like to work with people who share this view: engineers who enjoy kernels and compilers, researchers who care about training and quantization, and developers who want a better option than waiting for GPUs. See open roles, or follow our work on X and LinkedIn.