Frontier intelligence.
Small model footprint.
We post-train both frontier and small open-source LLMs on your data. By distilling knowledge from the best frontier model into a small OSS model, you get frontier-level intelligence on your task at a fraction of the compute cost — without compromising performance.
Frontier models are the most capable but the most expensive to run at scale — and your data has to leave your infrastructure to reach them. Small open-source models are cheap and fast, but out of the box they fall short on your task.
Distillation transfers what the frontier model has learned into a small model sized for your task, then post-training on your data closes the rest of the gap — so you keep frontier-level intelligence at small-model cost and latency.
From your data to a deployed model.
Bring your data (and harness, if you have one)
Send us your task data — labeled examples, transcripts, tickets, whatever defines the job. No labeled data yet? Pull from our Labeled Data and Feature Store pipelines instead. Already have an eval harness? Use it — if not, we'll build one with you so "better" is measurable from day one.
Spot-check frontier and small models
Before committing to a training run, we spot-check a range of frontier and small open-source models against your harness — so we know which models are already close, and which need real work.
Distill the best frontier model into a small OSS model
We transfer the reasoning and judgment of the strongest frontier model into a small open-source model — capturing the behavior, not just the outputs, so the small model inherits frontier-level intelligence.
Post-train and validate on your data
The distilled small model is post-trained on your data and re-run through your harness until it matches (or beats) frontier performance on your task — then it's ready to ship.
Not a separate product. The LLM layer of PreIntel.
Post-training runs on the same infrastructure as our Labeled Data, Feature Store, and Model Garden — it's the fourth layer, not a bolt-on.
No labeled examples for your task? Draw from the same human-verified datasets that train our ML/DL models.
Enrich your training and eval data with WHOIS, PDNS, cert, and content features via the same API.
Need a fast classical detector instead of an LLM? The Model Garden covers ML/DL threat detection.
Save on compute. Not on performance.
| Approach | Accuracy on your task | Cost | Latency | Data control |
|---|---|---|---|---|
| Frontier LLM API | Highest out of the box | $$$ per token, scales with volume | Higher, network round-trip | Third-party hosted, your data leaves your infra |
| Off-the-shelf small OSS model | Underperforms on your task | $ — cheap to run | Low | Self-hosted, full control |
| PreIntel distilled + post-trained modelRecommended | Matches frontier on your task | $ — small model economics | Low | Self-hosted or via our API |
Your data stays yours.
Your task data and any harness you provide are used only to evaluate and train your model. Nothing is shared with third parties or other customers.
If you don't have an evaluation harness, we work with you to define what "good" looks like for your task, then build the harness used to spot-check and validate every model.
Run the finished model via our hosted API, export the weights to your own infrastructure, or keep training in-house — the distilled model is yours.
Bring your data. We'll bring the frontier intelligence.
Post-training engagements are scoped per task — talk to us about your data and target model, and we'll spot-check candidates before you commit to a build.