MINIMAAI research lab Discuss fine-tuning

Fine-tuning + optimisation

Fit the model to your business. Then make it efficient.

You bring the fine-tuning data. Minima adapts the model to your task, then optimises the resulting model and delivers it with the runtime for your infrastructure.

Discuss your model and data Explore optimisation only

Explicit prerequisite

Your data is the starting point.

Task, dataset requirements and evaluation criteria are agreed before work starts.

CUSTOMER DATA

You provide the fine-tuning data.

Minima does not supply missing proprietary training data.

SCOPE

Agree the task and dataset requirements.

We establish what the model must learn and whether the available data is suitable before work starts.

EVALUATION

Define success before training.

Evaluation criteria and held-out evaluation data are agreed separately from the training set.

Workflow

Adapt first. Optimise second. Validate both.

  1. 01

    Scope

    Agree the task, base model, deployment hardware and success criteria.

  2. 02

    Prepare

    Review data suitability and separate training data from held-out evaluation sets.

  3. 03

    Fine-tune

    Adapt the model to the customer's task.

  4. 04

    Optimise

    Compress and prepare the fine-tuned model for efficient inference.

  5. 05

    Validate and deploy

    Evaluate task quality, serving performance and infrastructure requirements, then deliver the agreed weights and runtime.

Evaluation

Measure adaptation and optimisation separately.

Fine-tuning quality improvements are not attributed to compression.

Fine-tuning comparison

Task benefit versus the starting model

Evaluate the adapted model against the starting model using the agreed task and held-out evaluation criteria.

Optimisation comparison

Optimisation benefit versus the fine-tuned model

Evaluate the optimised model against the fine-tuned model before optimisation, including task quality and serving performance.

What you receive

The adapted deployment package.

  1. 01

    Adapted and optimised model.

  2. 02

    Agreed runtime.

  3. 03

    Evaluation findings.

  4. 04

    Deployment guidance.

Optimisation capability

Qwen3.8-2.4T

Four NVIDIA B200 nodes → one NVIDIA B200 node

Four NVIDIA B200 nodes for the baseline become one NVIDIA B200 node with Minima.

4× smallerModel weights
3.5× smallerKV cache
2× higherTokens per second

This is an optimisation result, not a fine-tuning experiment. It demonstrates Minima's optimisation capability. Achievable results for a customer's fine-tuned model require its own evaluation.

Commercial model

Fine-tuning and optimisation are scoped around your model and data.

Production deployment uses an annual software licence priced on agreed GPU-hour usage.

Fine-tuning FAQ

Data, evaluation and deployment.

What data is required?

You provide the fine-tuning data. We agree the task and dataset requirements, review data suitability, and separate training data from held-out evaluation data before work starts.

Can we use an already fine-tuned model?

Yes. Minima can evaluate an already fine-tuned model as the starting point, subject to access to the necessary model artefacts.

How is quality measured?

Fine-tuning benefit is measured against the starting model using agreed task criteria. Optimisation is evaluated separately against the fine-tuned model before optimisation.

Is optimisation-only available?

Yes. Explore the Enterprise optimisation offer for optimisation and deployment without fine-tuning.

Fine-tuning + optimisation

Discuss fine-tuning

Tell us about your task, model, data readiness and deployment environment so Minima can scope the fine-tuning and optimisation work.

By submitting, you ask Minima to contact you about fine-tuning and optimisation. See our Privacy Policy and Terms of Service.