Inception Labs
A lab building Mercury, a language model that generates text by diffusion instead of one token at a time.
Highlights
- Mercury — diffusion-based large language model architecture (vs autoregressive transformers)
- Parallel token generation produces ~10× faster inference at comparable quality
- Founded 2024 by Stefano Ermon (Stanford CS faculty, diffusion-model theorist)
- Approximately $20M+ seed; Khosla Ventures lead with Mayfield and M12 participation
- API access available with competitive pricing vs OpenAI / Anthropic
- Stanford spinout with strong academic research lineage
- Targets latency-sensitive applications where autoregressive models hit speed ceilings
External link — opens inceptionlabs.ai in a new tab. Inception Labs is a third-party product; we are not affiliated with it.
About Inception Labs
What it is
Inception Labs is a Stanford spinout building Mercury, a family of diffusion-based large language models. Where a conventional transformer produces text sequentially, one token conditioned on the last, a diffusion model refines the whole output in parallel — which the company reports gives roughly ten times faster inference at comparable quality. Founded in 2024 by Stefano Ermon, a Stanford faculty member and diffusion-model theorist, with Khosla Ventures leading the seed. API access is available.
Why it's different
Nearly every model you have used is autoregressive, so this is a genuine architectural departure rather than a tuning difference, and the speed claim follows directly from the maths rather than from better hardware. That makes it interesting for anything latency-bound. The caveats are proportionate to the novelty: the quality comparisons are largely the company's own, the ecosystem of tooling and fine-tuning support that exists for transformers does not exist here, and betting a product on a single small lab's architecture is a different risk from using an established provider. Promising is the right word, not proven.
How people use it
The use that makes sense today is anything where response latency is the constraint and absolute quality is not — autocomplete, live assistance, interactive agents where a two-second pause breaks the interaction. Developers typically benchmark it against their existing provider on their own prompts rather than trusting published figures, because the speed advantage is real but the quality trade varies a lot by task.
Written by the n3os team. We are not affiliated with Inception Labs.
This listing was written from public information, without Inception Labs’s involvement. If you own it and something here is wrong — or you would rather not be listed at all — email us and we will correct or remove it.
Get the ones worth knowing about
We write one of these for every tool worth the trouble. Get the new ones, plus what we have found genuinely useful lately.
Your address goes to Buttondown, who send the emails on our behalf. One click unsubscribes, and the list is never sold or shared.