kolm is the compiler layer above device runtimes.
Core ML, LiteRT, ONNX Runtime, and ExecuTorch are excellent runtimes. kolm sits one layer above them: it compiles regulated workflows into signed artifacts with receipts, K-score gates, and a compliance trail those runtimes do not own.
where kolm fitsPrivacy and compliance are not runtime features by themselves. They need artifact provenance, eval evidence, and operational receipts.
Do not fight the runtimes. Compile for them.
The serious competitors are free, mature, and backed by Apple, Google, Microsoft, Meta, and the PyTorch ecosystem. The credible kolm strategy is to sit above them as the workflow compiler, quality gate, and trust layer.
| System | Primary layer | Strongest use case | kolm position |
|---|---|---|---|
| Apple Core ML | Apple on-device model runtime | iOS/macOS apps that need deep Apple silicon integration. | kolm can target Core ML for Apple deployments while adding signed workflow evidence above the model file. |
| Google LiteRT | Cross-platform on-device runtime | Android, iOS, desktop, web, IoT, and micro deployment with accelerator support. | kolm can emit LiteRT-targeted artifacts for teams that need a regulated compile process and receipts. |
| ONNX Runtime | Cross-framework model accelerator | Teams standardizing on ONNX across frameworks and hardware execution providers. | kolm can package ONNX Runtime as an execution target instead of inventing another generic accelerator. |
| ExecuTorch | PyTorch edge deployment stack | PyTorch teams exporting models to mobile, embedded, and edge hardware. | kolm can make PyTorch deployment auditable by bundling evals, manifests, and receipts around exported targets. |
| kolm | Compiler, artifact, registry, verifier, and receipt layer | Regulated product teams that must prove what was compiled, what passed, what ran, and where data went. | kolm should orchestrate runtimes, not replace them. |
kolm vs Predibase, OpenPipe, Fireworks, Together.
The closest substitutes are the platforms that promise "fine-tuning made easy" or "your own model on our infrastructure." They emit a hosted endpoint. kolm emits a signed file you own. The two are usable together, but only one compounds when you stop paying.
| Vendor | What you get | Where it runs | Lock-in | kolm position |
|---|---|---|---|---|
| Predibase | LoRA / adapter training on managed serverless infra; API endpoint per fine-tune. | Predibase cloud only. | Adapters portable; runtime inference is theirs. | kolm packages the same LoRA into a runnable .kolm with eval, K-score, and receipts. You can keep Predibase as a training back-end and ship a portable artifact. |
| OpenPipe | Production capture → auto-fine-tune → hosted inference; cost-reduction wedge. | OpenPipe cloud (or BYO inference). | The pipeline lives on OpenPipe. Pulling out is a project. | kolm is the artifact pipeline. Capture → verifier → LoRA → signed file. The artifact moves with you; the registry replaces the platform graph. |
| Fireworks AI | Fast inference for open weights, with fine-tuning in beta. The throughput-per-dollar play. | Fireworks cloud. | Endpoint stays on their infra. No artifact you can mail. | kolm is offline-first. The same compiled artifact runs on Fireworks (hot path) and on a phone (cold/private path). Fireworks becomes one optional run target, not the only one. |
| Together AI | Open-model serving + fine-tune; OSS-friendly, broad model menu. | Together cloud, with private deploy contracts. | Weights are theoretically portable; the running config and eval pipeline are not. | kolm packages weights + adapter + recipe pack + verifier + eval suite + receipts in one zip. The eval and the gate travel with the model; that's what makes it auditable, not just hostable. |
| kolm | A signed .kolm file: model + LoRA + recipe pack + recall index + eval suite + manifest + receipts. |
Anywhere: laptop, phone, server, edge box, or any of the above platforms. | None. The artifact is yours; the registry is open (RS-1, MIT). | kolm is the compiler. The output is the product, not the endpoint. |
What only kolm gives you
A single signed file. Byte-identical reproducibility. K-score gate before emit. HMAC-chained receipts at every inference. Offline runtime as a first-class target.
What the platforms still do better
Hosted GPU throughput at production scale, batched serving, vendor-managed uptime. If your only deploy target is a hosted API, a platform is a fine choice, but you do not own the artifact.
Useful combinations
Use Together or Fireworks as a hot inference target for the same .kolm you ship to phones. Use Predibase as a training back-end behind a kolm bridge. Same artifact, more places it runs.
The bar for winning is evidence.
A developer will not switch because a landing page says "private AI." They switch when kolm gives them a workflow they cannot get from free runtime tooling alone.
Artifact provenance.
Every `.kolm` needs a manifest, data-boundary declaration, eval set, K-score, hash chain, and reproducible build metadata.
Runtime neutrality.
Core ML, LiteRT, ONNX Runtime, and ExecuTorch should be deploy targets. The compiler owns policy and evidence.
Regulated workflow fit.
HIPAA, fintech, and enterprise mobile teams need auditable local inference, not another model conversion CLI.
Benchmarks.
Latency, size, accuracy, battery, and privacy evidence must be published per hardware target, not implied.
Use the runtime directly when trust evidence is not the bottleneck.
If you already have a model, a single platform, no regulated data, and no need for receipts, use the native runtime directly. kolm is for teams whose blocker is shipping a private AI workflow through engineering, security, legal, and procurement.
Use Core ML directly
When the target is Apple-only and your team already owns model conversion, testing, packaging, and release review.
Use LiteRT directly
When you need flexible cross-platform inference but do not need signed compile receipts or compliance artifacts.
Use ExecuTorch directly
When your ML team lives in PyTorch and wants the shortest export path to edge devices.
Official references.
Every claim above is from the vendor's own positioning. Where a feature is in beta or behind contract, we say so; we do not editorialize.
Predibase
OpenPipe
Fireworks AI
Together AI
Apple Core ML
Google LiteRT
ONNX Runtime
ExecuTorch
Ship private AI as an artifact, not a promise.
Start with the compiler. Prove the artifact. Then dispatch to the best runtime for each device.