The model gets smarter.
Better reasoning on harder problems. As more diverse usage arrives, the next-token prediction sharpens toward SOTA-class on the eval suite.
The flywheel. The privacy contract. The data path.
Three things to understand about Tessera: what the flywheel does, why your data is safe, and what each release proves. In plain language. The technical depth lives on the architecture page.
The flywheel
It gets better the more people use it.
Tessera is not just an inference engine. It is a system that improves the model through collective usage. The pattern is straightforward: a user runs a model on their device, the system observes a sanitized baseline of that usage, the observation ships to the Calibration Commons, the next Tessera release carries the improvement, and every Tessera user gets the new model. The user's local experience drives the global improvement; the global improvement is delivered to the user's local experience.
What improves, in concrete terms
The hybrid drafters get more accurate.
The drafter proposes tokens; the verifier accepts or rejects. Higher acceptance rate means the same quality with fewer forward passes. Faster and more accurate, not just one or the other.
The engine / harness gets faster.
The dispatch policy learns to put the right operation on the right silicon. Quantization-on-transport learns the access patterns the model actually exhibits.
The next-token prediction sharpens.
As the training distribution grows, the model's per-token loss on held-out evaluation sets drops. The flywheel's primary observable is the eval deltas on each release.
The privacy contract
Raw data never leaves. The audit trail is public.
Tessera is built on a strict separation between what is public and what is private. The privacy contract is short, and it is enforced by the design, not by promises.
Public: the contract
What the user agrees to, what the system is allowed to do, what the audit trail proves. The license, the receipt schema, the eval results, the snapshot fingerprints. Anyone can read this; no NDA required.
Public: the audit trail
Every Tessera release ships with the recipe that produced it: the upstream model SHA, the calibration corpus snapshot, the per-tensor policy, the eval results, and a signed receipt. Anyone can verify what the model is. The method by which the data was sanitized is private; the proof that it was sanitized is public.
Private: the sanitization method
The pipeline that turns raw usage signals into license-clean, anonymized training material. The method is the proprietary data work that makes the trained models defensible. Public, the method could be copied; private, the auditability stands on its own.
Private: the raw inputs
Your device's raw usage data — the actual text, audio, or video that flows through Tessera Studio — never leaves the device. The sanitization runs on the device; only the sanitized output is eligible to ship, and only with your opt-in. There is no cloud collection step.
The same posture governs the S2S (speech-to-speech) consent lane. The contract and the auditability are public. The method is not. S2S design doc for the precedent.
The data path
What flows where, step by step.
The flywheel is opt-in. With opt-in on, the path is the same every time:
-
01
Local usage
You use Tessera Studio. Your text, vision, audio, and speech prompts run on your Mac. The model answers. Nothing leaves yet.
-
02
Sanitization on device
During idle time, Tessera Studio runs the sanitization pipeline locally. The pipeline transforms the recent usage into a license-clean, anonymized baseline. The raw inputs are discarded; the sanitized output is the only thing eligible to ship.
-
03
Consent, then ship
You see a one-line summary of what would be shipped. You can read the sanitized output, edit it, or skip the upload. With your consent, the sanitized output is added to the Calibration Commons with a snapshot fingerprint.
-
04
Next release
The next Tessera model release includes the contributions from the most recent Commons snapshot. Every Tessera user gets the new model. The signed receipt on the new model includes the Commons snapshot fingerprint, so you can verify which contributions went into the model you are using.
Reciprocity
The community tier is not a charity.
Tessera Studio is free. Tessera-trained models are CC BY-NC-SA 4.0. The engine is source-available under the Tessera Research and Education License 1.0. None of that is philanthropy. It is the front door of a relationship.
When you use Tessera Studio, you are part of the flywheel. The flywheel is what makes the model get smarter. The model getting smarter is what makes Tessera worth using. The relationship is structural, not transactional: you do not need to "join the Commons" or "contribute compute" by name. You just use the app, opt in to the flywheel, and the rest happens.
Commercial use is a separate path. julian@tribunus.dev for a commercial license. The commercial licensee gets the right to redistribute under their own terms; the community tier does not change.
Where to go next