kolm / compare / vs OpenAI fine-tune
kolm vs OpenAI fine-tune.
OpenAI fine-tune produces a closed-weight model that lives inside OpenAI's cloud and bills per token. The kolm artifact is an open-weight signed file on your disk that bills per compile. The deeper difference is what you can audit afterwards.
Ten axes. Reviewed 2026-05-15.
| Axis | kolm | OpenAI fine-tune | Why it matters | Proof |
|---|---|---|---|---|
| Weight access | open, you own the file | closed, OpenAI keeps the weights | If you do not have the weights you cannot run, audit, or migrate. | anatomy → |
| Run location | anywhere your hardware is | OpenAI cloud only | Regulated data and edge devices cannot leave the perimeter. | /edge → |
| Cost | flat compile, $0 to serve | per-token forever, premium rate | Fine-tuned token pricing is roughly 8× the base. The line crosses at modest volume. | /roi → |
| Audit | receipts + manifest + binder | opaque | If your auditor asks "what did the model see and what did it return on this date" you need a primitive. | receipt JSON → |
| Algorithms | 42 SOTA (DPO, KTO, ORPO, SimPO, GRPO, distillation, MoE-LoRA, federated) | SFT only | Different task shapes need different alignment recipes. | /research → |
| Replay forever | yes, byte-deterministic | no | A regulator may ask you to replay a 2024 decision in 2027. | RS-1 → |
| Compliance | BAA template + on-prem + HIPAA pack | enterprise DPA, shared infra | Healthcare and finance procurement walks if data leaves the perimeter at run time. | BAA → |
| Quantization | INT4 / INT8 / NVFP4 baked at compile | N/A | If the model is in their cloud they pick the precision. | NVFP4 → |
| K-score quality gate | 0.40·A + 0.15·S + 0.15·L + 0.15·C + 0.15·V | accuracy in dashboard | A blocking gate that ships in the receipt is what the procurement officer reads. | /k-score → |
| Multi-tenant residency | on-prem per tenant | OpenAI shared | Each tenant's data and model in its own perimeter. | /airgap → |
When OpenAI fine-tune is the right answer.
You want frontier weights as a starting point, you do not need to run anywhere except their API, and the cost line will not cross. For a small consumer product calling a fine-tuned 4o once a session, this is the path of least resistance.
When kolm is the right answer.
You operate in a regulated industry, you need to own the weights, you want audit primitives that survive a 2027 audit of a 2024 decision, or your token bill is already structural. The signed file is what every other tier of the company eventually asks for.