The AI compiler.
Give kolm your data and a task. Get back a single file that runs anywhere, even offline. The model is yours.
One command.
Three lines of context.
A model that's yours.
Four engines. One command.
A compile is the choreography of recall, distill, decompose, and run. You see one line. We do the rest.
Index your world.
Text, image, audio, PDF, video. Embedded once. Queried by every compile.
Borrow the frontier.
A frontier model labels the task. Every output verified before it counts.
Cache the patterns.
Deterministic outputs become a draft pack. Speculative decoding for free.
Ship the artifact.
Model, LoRA, draft pack, index. Composed at every token. Offline.
Use it three ways.
Same compiler. Same artifact. Three doors in.
Terminal first.
For developers. Pipe a folder. Pipe a database. Get back a .kolm.
$ npm i -g @kolmogorov/kolm $ kolm compile "spot bugs like me" > bugs.kolm readyInstall the CLI ›
Browser. Drag. Click.
For teams. Drop examples. Drop a corpus. Click compile. Download or call by HTTPS.
https://kolm.ai/compile kolm.compile(task).run(input)Open the cloud ›
Pocket size.
For everyone. Three Specialists out of the box. Compile a new one in a tap.
iOS · Android · PWA tap Compile.See the app ›
Inside a .kolm file.
A signed zip. Sub three gigabytes. Boots on a phone.
- model.ggufBase model. INT4 quantized. Boots on a 2021 phone.
- lora.binYour task fine-tune. Trained on verified labels.
- recipes.jsonDeterministic draft pack. Speculative decoding.
- index.sqlite-vecMultimodal embeddings of your corpus.
- manifest.jsonStats, hashes, registry version.
- signature.sigHMAC chain anchored to a public registry.
Frontier inside. Yours outside.
A new compute compact. Borrow the frontier to teach. Own the result.
Frontier teaches. You ship.
Distill against any frontier model with your own key. The student fits in your pocket. The teacher stays at home.
Sub three gigabytes.
Quantized base, sparse LoRA, deterministic draft pack. Loads on any phone shipped since 2021.
Signed end to end.
Every artifact carries an HMAC chain anchored to a public registry. Every output verifiable.
Open spec. Open registry.
Recipe, Recall, and Receipts ship as RS-1 specs under MIT. The compiler is ours. The format is everyone's.
Receipts, not promises.
Every compile and every run produces a signed receipt. Inputs hashed, model versions pinned, decisions recorded. Verifiable by anyone, anywhere.
One product. Three tiers.
Pay for compute and storage. Run is always free.
- 1 Specialist a month
- 10k corpus rows
- 4B-class open base
- 100k managed vectors
- 1 personal Specialist
- 100k corpus rows
- 1M managed vectors
- Unlimited on-device
- Unlimited Specialists
- 1M rows each
- 10M managed vectors
- All open bases incl. Hermes
- Private base models
- On-prem trainer bridge
- Dedicated Trieve cluster
- On-chain receipts