Fixed quotes. Compute at cost. Your weights.
Clear packages for the most common engagements. Every quote is fixed against a written success metric, and GPU time is passed through with zero markup.
Feasibility Sprint
Answer “will fine-tuning work for this task, and by how much?” before committing a budget.
- Data audit & labelling plan
- Baseline eval suite (yours to keep)
- Prototype checkpoint on your data
- Go / no-go report with a fixed quote
Production Model
One task, taken from data to a model serving production traffic, with the evidence to sign it off.
- Data program incl. expert annotation
- Base-model bake-off, SFT + preference tuning
- Held-out evals, red-team, regression gates
- Quantized deployment in your cloud
- 60 days of post-launch support
Model Program
A dedicated pod running a portfolio of models, with continuous retraining and new use cases each quarter.
- Research lead + 2 engineers
- Monthly retraining on fresh data
- Drift monitoring & on-call
- New model every quarter
Enterprise
Air-gapped and on-prem delivery, frontier-scale models, multi-team rollouts and bespoke security terms.
- On-prem & sovereign deployments
- 70B–400B+ parameter models
- Named senior researchers
- Custom MSA, DPA & security review
Prices in USD, excluding GPU compute (billed at cost) and applicable taxes. Production Model scope assumes a single task and up to 20,000 expert labels.
Build in-house, buy a platform, or hire a lab?
An honest comparison. Platforms are excellent if you already have a post-training team. Most companies don’t — yet.
| Hire & build in-house | Self-serve fine-tuning API | Finetune the World | |
|---|---|---|---|
| Time to first production model | 6–12 months (hiring included) | Weeks, if your team knows what to do | 6.5 weeks median |
| Year-one cost | $1.2M+ for a 3-person team plus compute | Low tool cost, high internal time | Fixed quote + compute at cost |
| Data strategy & annotation | You build it | Not included | Included, with expert raters |
| Preference tuning & RL | Requires specialist hires | Usually SFT/DPO only | DPO, KTO, GRPO, PPO, verifiers |
| Evaluation harness | You build it | Generic metrics | Task-specific, yours to keep |
| Runs inside your VPC / on-prem | Yes | Often not | Yes |
| Weights ownership & portability | Yes | Varies by vendor | Always yours |
Four variables decide most of the bill.
Model size matters less than people expect. Data and evaluation are where projects are won or lost, so that is where the budget goes. Our fine-tuning cost guide has worked examples.
Pricing questions
Why fixed price instead of time and materials?
Because the risk should sit with the people who know how long the work takes. We scope against a written success metric. If we miss the timeline for reasons within our control, the overrun is on us.
How is GPU compute billed?
At cost, with no markup, either directly on your cloud account or invoiced from our reserved capacity. A typical 8B production model uses $3,000–$15,000 of compute across all experiments; 70B-class models run $20,000–$90,000.
What if the feasibility sprint shows fine-tuning won’t work?
You get a written recommendation (often better retrieval, a different base model, or a workflow change), the baseline eval suite and the data audit. Roughly one in five sprints ends that way, and those clients frequently come back with the right problem.
Do you offer annual or multi-model agreements?
Yes. Model Program retainers cover a dedicated pod and a pipeline of models with quarterly retraining. Enterprise agreements add on-prem delivery, named researchers and custom security terms.
Are there per-token or per-seat fees after delivery?
No. You own the weights and the code. The only ongoing cost is optional: a run-and-retrain retainer if you want us to operate the model.
Get a fixed quote in one call.
Tell us the task, the volume and where it must run. We come back within three business days with a scoped, fixed-price proposal.
Book a technical briefing →