kolmogorov cloud · the hosted compile + verified wrap

Bring your own key. Pay for what compiles.

The cloud runs the four engines so your laptop doesn't have to. Same compile pipeline, same artifact format, same K-score, plus a verified-inference wrap that turns any Anthropic / OpenAI / Mistral key into a self-checking call, k-sample, deterministically verify, accept the winner, return a receipt.

Four engines, one pipeline.

Cloud runs the four engines so your laptop doesn't have to. Each is a stateless service behind one HTTP boundary; the orchestrator schedules them per compile job. Everything is observable in /account.

RECALL

multimodal index

BM25 + vector + RRF + cross-encoder rerank. BGE-M3 1024d INT8. Per-tenant namespace.

DISTILL

verified inference

k-sample frontier model, deterministically verify, accept the winner, return a signed receipt.

DECOMPOSE

recipe drafts

Extract the structured-token subset of a model's behavior into a deterministic, registry-indexed draft pack.

RUN

artifact assembly

Bundle base GGUF + LoRA + recipes + sqlite-vec index + manifest, sign with HMAC chain, ship a single .kolm.

Measured latency, west-region

recall query~18 msp50, 100k chunks warm
verified wrap (k=3)~2.4 sp50 cold, ~1.1 s warm
compile job (5k rows)~3.4 minp50 end-to-end
artifact download~280 msp50, 17 MB .kolm
compile throughput~7.4 / minconcurrency 8, sustained

Verified inference, drop-in.

# the SDK is a transparent Proxy, wrap any Anthropic-shaped client import Anthropic from '@anthropic-ai/sdk'; import { kolm } from 'https://kolm.ai/sdk.js'; const raw = new Anthropic({ apiKey: 'sk-ant-...' }); const verified = kolm.wrap(raw, { verified: true, k: 3, corpus_namespace: 'work' }); // same shape as the official SDK; routed through /v1/wrap/verified const r = await verified.messages.create({ model: 'claude-opus-4-7', max_tokens: 1024, messages: [{ role: 'user', content: 'classify intent: ...' }], }); console.log(r.content); // the winner console.log(r._kolm.k_samples); // 3 console.log(r._kolm.verifier_pass); // true console.log(r._kolm.receipt); // signed proof

Or just compile in the browser.

Drag a folder onto the Cloud dashboard, write a task in the textbox, click Compile. The pipeline runs on our hardware and emits a downloadable .kolm with full K-score breakdown. Your data is namespaced per tenant, your API key is yours, your model bill stays on your account.

# or via curl, if you live there $ curl -X POST https://kolm.ai/v1/compile \ -H "Authorization: Bearer $KOLM_KEY" \ -H "Content-Type: application/json" \ -d '{ "task": "classify support tickets", "examples": [...], "corpus_namespace": "work" }' { "job_id": "job_abc", "status": "queued" } $ curl https://kolm.ai/v1/compile/job_abc/.kolm -o triage.kolm

Pricing.

Free

$0 / mo
  • 1 Specialist / mo
  • 10k corpus rows
  • 4B-class open base only
  • 100k vectors managed
  • unlimited kolm run
Start →

Mobile

$9 / mo
  • 1 personal Specialist / mo
  • 100k corpus rows
  • 1M vectors managed
  • kolm.app phone runtime
  • unlimited on-device
Start →

Pro

$49 / mo
  • unlimited Specialists
  • 1M rows each
  • all open bases incl. Hermes-3
  • 10M vectors managed
  • full CLI + Cloud + MCP
Start →

Enterprise

custom
  • private base models
  • on-prem training bridge
  • private Trieve cluster
  • dedicated GPU embedders
  • on-chain receipt anchoring
Contact ›

You always pay your own frontier API bill on top, we're the compiler, not the model host.