A practical specialist-model system
Work with us.
Keep control.
We help teams decide whether a small model is justified, then build and measure the smallest system that meets the task. If a frontier model already clears your bar, we’ll say so.
One job, clearly defined.
We agree on inputs, expected outputs, failure cost, quality target, latency, privacy, and target devices before training starts.
Your data stays yours.
We work from data you have the right to use. Training and held-out evaluation examples stay separate so a good result means something.
Run it without us.
You receive model artifacts and license notes, evaluation code, device measurements, integration guidance, and operating documentation.
1–2 weeks for a fit assessment and baseline report. 3–6 weeks for data preparation, adaptation, evaluation, and packaging. Integration time depends on your product team.
One narrow task, 50–500 representative examples you can authorize, the target device or server, and one person who can approve data use and quality thresholds.
A written fit decision, a measured baseline, the adapted checkpoint if justified, evaluation scripts, device numbers, and a handover runbook.
Send one task with a redacted input and the output you expect. We reply with a short fit note and the first question that decides the project.
One project.
Clear gates.
Each phase ends with an agreed output or a decision to stop. You stay involved in the quality, data, and deployment choices.
- 01First contact and fitOutput · fit summary
Share one narrow task, a rough example, and the constraints you already know. We identify a useful first question and flag disqualifiers.
- 02Scope and success criteriaOutput · project brief
Agree on baseline systems, quality thresholds, latency, memory, privacy, license, ownership, and what counts as a successful deployment.
- 03Prepare and approve dataGate · data approved
Review data rights and handling, redact sensitive fields where needed, build examples, and reserve a held-out evaluation set.
- 04Evaluate the frontier and small modelsGate · baseline decision
Score the current best general model and a shortlist of open-weight specialists on the same representative cases. Stop if the baseline is sufficient.
- 05Adapt, test, and packageGate · agreed eval and fit
Use retrieval, prompting, fine-tuning, or distillation only where results justify it. Quantize and profile the deployment on the target hardware.
- 06Deploy and hand overOutput · operable system
Ship to the agreed device or private server. Hand over weights, evaluation scripts, runbooks, monitoring signals, and a maintenance plan.
Phases 03, 04, and 05 are gates. A failed gate ends the project early, and you keep the evaluation and the written decision. That is a valid result, not a wasted phase.
See a redacted sample: fit summary and baseline result
A real deliverable is longer than this. The shape is the same: a decision, the evidence, and the next action.
| Item | Example | Decision |
|---|---|---|
| Task | Extract 6 fields from supplier invoices (PDF and scan) | In scope |
| Baseline | Frontier model, 94% field accuracy on 120 held-out invoices | Clears the 90% bar |
| Candidate | Qwen3.5-0.8B, 4-bit, 8K context, laptop CPU | Fits the 8 GB budget |
| Result | Fine-tuned candidate: 92% field accuracy, 11 tok/s | Pass · deploy candidate |
| Trade-off | Frontier model is 2 points better but sends data off-device | Client chooses on-device |
Figures are illustrative. Every project gets its own baseline and held-out set. We do not reuse a result from another client.
No deck needed
Tell us what the model must do.
The form opens your email app. You can also write to hello@talachu.com. Prefer a call? Email us to arrange a 20-minute fit call. Typical work starts with a 1–2 week evaluation phase; timelines depend on data readiness and integration.
Data, ownership, and licenses
How is client data handled?
Handling and retention are agreed before data transfer. Share redacted examples for initial scoping. Training data is separated from held-out evaluation data.
Who owns adapted weights?
Ownership and delivery terms are written into the project scope before work begins. You keep your data and your evaluation examples.
Can every model be used commercially?
No. Each checkpoint has its own license and use restrictions; we review the exact terms before selecting it. The model index shows a summary, not legal advice.
What if the small model does not beat the baseline?
We say so in the baseline report. You keep the evaluation, the measured gap, and the written recommendation. No adaptation work follows without your agreement.
Do you take small projects?
Yes, if the task is narrow and you can provide representative examples. A fit assessment is the smallest engagement.
Will you sign an NDA?
Yes, before detailed data is shared. Start with a redacted example so we can decide whether the project is a fit.
Do you resell models?
No. We deliver work under the license of the underlying checkpoint and the terms in the project scope.