Your model, distilled.
A fraction of the cost.
Turn an expensive production LLM task into a fast, specialized small model, trained for you and deployed on dedicated GPUs in Europe. Same accuracy, up to 80% lower inference cost, with just one endpoint to swap.
distil labs distil labs benchmark
Two specialists, one endpoint
distil labs builds the model. Lyceum runs it on European infrastructure. You get the accuracy of a large model at the cost and speed of a small one.
- Curates data from your traces
- Distills a specialized sub-8B model
- Evaluates against your acceptance criteria
- Dedicated GPU endpoint in the EU
- Autoscaling and operations handled
- GDPR-compliant by default
Accurate or affordable. You shouldn't have to choose.
At scale, inference becomes a real line item, and every option forces a trade-off.
Frontier models are expensive
Large proprietary models are accurate, but at hundreds of millions of requests a month even the smallest variants add up fast.
Open-source rarely matches both
Off-the-shelf open models tend to match proprietary ones on accuracy or on efficiency, but not on both at once.
Fine-tuning needs a team
Training data, ML expertise, GPUs and an endpoint to run it. Most teams stall before the ROI is ever proven.
From production trace to live endpoint in days
We distill a large open teacher model into a small model specialized to your task, then host it for you.
Share a system prompt & traces
Send the system prompt for the task and roughly a day of production traces. No data-prep or labelling work on your side.
distil labs trains the model
A large open teacher model is distilled into a specialized sub-8B model, tuned to your task and evaluated against your acceptance criteria.
Lyceum deploys it in Europe
We host the model on dedicated GPUs in European data centers, with autoscaling handled for you. GDPR-compliant by default.
Swap one API endpoint
Point your existing workflow at the new endpoint. No other code changes. We manage the endpoint and keep it healthy.
distil labs case study 68% lower inference cost, no drop in quality
This is a proven result from distil labs, not from us. Their knowledge distillation cut inference cost by 68% for a leading European edtech platform serving tens of millions of students, at higher accuracy, across hundreds of millions of requests a month. Lyceum's role in the partnership is to run models like this on dedicated European infrastructure.
| Proprietary small model | Custom distilled model | |
|---|---|---|
| Classification accuracy | 81% | 93% |
| Latency p50 / p95 / p99 | 0.49 / 0.81 / 1.28 s | 0.27 / 0.59 / 0.66 s |
| Cost per million requests | $109 | $35 |
Benchmark and results reported by distil labs, measured on held-out production traffic. Pricing is based on instance uptime from real production usage over several days, not synthetic benchmarks.
Why teams distill with us
The full custom-model lifecycle, training, deployment and endpoint operation, handled for you.
Days, not quarters
A first production-ready model, trained, deployed and ready for testing in 3-5 days. Not a hiring cycle.
No ML team required
Model training, deployment and endpoint optimization are fully managed. You just exchange the API endpoint.
Accuracy held or improved
Because the model is specialized to one task, quality matches, and often beats, the large model it replaces.
Up to ~80% cheaper inference
A small, fast model on dedicated GPUs runs at a fraction of the cost of frontier or proprietary small models.
EU data residency
Weights are hosted on Lyceum infrastructure in European data centers. Your prompts never transit US infrastructure.
Data-efficient distillation
Knowledge distillation needs fewer than 100 examples to get started. No large labelled dataset required.
Have a task that's too expensive to run at scale?
Tell us about your use case and current model. distil labs will scope a custom distilled model with you, free to train, and Lyceum will have an endpoint ready to test in days.