We build models that dream first.
Tannic Learning Models carry a world model inside them. While training they rehearse futures. At inference they score every option and commit to the best one.
Latent dream sphere
We build models that dream first.
Tannic Learning Models carry a world model inside them. While training they rehearse futures. At inference they score every option and commit to the best one.
Tannic ML is an AI research lab. We build language models that don't just answer, they imagine. Our work lives where machine learning meets world models, and it comes down to one idea: a model that can dream learns faster and decides better.
We train on open models, we publish what we build, and we keep the work small and legible.
TLM stands for Tannic Learning Models. We start from strong open models, LFM2.5-8B-A1B and Gemma 4 12B, then train them with techniques we invented around world models.
What comes out is a family of fast, capable models that carry a world model in their wiring. Not an add-on. Part of how they learn and how they think.
Dreaming is our core technique. At training time the model rehearses imagined futures and learns from them, which speeds up learning. At inference time it drafts several continuations, scores each one against its world model, and commits to the one that scores best.
The backbone is an ordinary LLM, kept frozen under LoRA. A chunk pooler reads hidden states at response boundaries and collapses each span into a latent point. A predictor learns how those points move, and a critic learns which futures are worth keeping.
K short drafts share one cache prefix. Each draft is rolled forward in latent space, scored by the critic, and only the winner gets decoded.
Our models live on Hugging Face, with weights, world models, and notes on how each was trained. Start with the TLM-1 model collection.
TLM-1-Action is our first 4B model, built on Cosmos3-Edge. The backbone stays frozen under LoRA while the world model dreams over action trajectories instead of text: it drafts candidate actions, scores each one in latent space, and commits the winner.
Trained on robot manipulation data. Still under training.
Notes on what we're building: model releases, training runs, and the occasional essay on dreaming at scale.
Follow the model family on Hugging Face for releases as they land.
Questions about the models, the techniques, or a collaboration? Send us a note.
We read everything, and we answer most of it.
We're a small lab and we plan to stay small. We're not hiring right now, but when we open roles they'll show up here.
No open roles right now.