Skip to main content

Intelligence,
tile by tile.

Research

Latest from the lab

All research→
#efficient-reasoning

Tessera-27B-B-CJK: teaching a 27B agent to think less per turn

Tessera-27B-B-CJK is Qwen3.8-27B with a rank-32 LoRA trained on Tinker. It thinks in a compressed CJK shorthand that mixes English, CJK words and symbols, and we release it as an agentic model rather than a general reasoner. On 44 frozen states from 11 held-out coding-agent tasks it uses about 0.50-0.64x the base model's thinking tokens per agentic turn, from the typical turn to the token-weighted ratio. It made a valid tool call on all 88 of its draws, with no repetition loop in any of them, and on 6 long-context states with inputs of 101K-131K tokens it made a valid tool call 6 times out of 6. The safety smoke test raised no flag against the base. On hard math and competitive code it is worse than the base, and it loops more often there.

Read paper→

Approach

How we work

  1. 01

    Measured

    We compare every result against its baseline on the same budget, and report the gap either way.

  2. 02

    Deliberately open

    We publish our research and release open weights when a model is ready for the community. The training systems, data pipelines, and products built on that work are the lab's own.

  3. 03

    Experimental

    We try ideas the field has barely tested, like reasoning in a compressed notation or reusing a model's cached state in a later session. Many of them don't work.

Model lineupsIn development

Mosaic & Tessera

Models built from our research.

We're turning our research into models you can use. Mosaic is in development, and the first Tessera models are out.

Mosaic

Our foundation model lineup.

  • Mosaic ProThe larger model in the family.
  • Mosaic MicroThe compact model in the family.

Tessera

Existing models, tuned with our research, starting with reasoning compression.

Chat isn’t open yet. Join the waitlist and we’ll email you when sign-up opens.

Join the waitlist→

FAQ

Questions we get a lot

If something isn't answered here, reach out via contact.

  • An independent AI research lab. We publish our research, release open weights, and build the Mosaic and Tessera models from what we learn.

  • Our recent papers look at reasoning in fewer tokens, reusing a model's cached state across sessions, faster memory updates in memory-augmented transformers, and which layers can give up bits when weights are compressed.

  • Research papers and technical reports, open weights for selected models, and occasionally code. Not everything: training infrastructure, data pipelines, and our products stay in-house. We share the tiles; the pattern is ours. What we do publish includes the experiments that don't work.

  • Yes, the Tessera models we've released. Their weights are on Hugging Face, and each model's paper covers what it does well and where it falls short. Mosaic Pro and Mosaic Micro are still in development; we plan to release their weights too. Chat access will follow.

  • We're smaller and research-first. The results we publish are the same ones our models are built from, including the experiments that fail, so anyone can check the work behind what we ship.

  • Read our research, follow us on X, or reach out if you want to collaborate. We're always interested in talking to people doing interesting work in this space.