Hosted fine-tune. Or compiled file.

Same training step, different deployment

Both products do supervised fine-tuning on open weight bases (Llama, Qwen, Mistral, etc.). The training step itself is similar - LoRA or full-parameter, configurable rank, configurable epochs. Where they diverge is what comes out.

Together returns a hosted model endpoint with a per-token rate sheet. kolm returns a .kolm file: weights, recipes, manifest, signature. From the moment you have the file, your inference cost is "what you pay your hardware vendor," not "what you pay Together per token."

Where Together wins

Honest concession. Together's serving infrastructure is excellent. If you don't want to operate inference servers, the GPU pool is mature, the autoscaling works, the rate-limiting is tight, and the OpenAI-compatible API is one drop-in line of code. For a team that just wants the model running and isn't ready to manage hardware, this is the fastest exit.

Model breadth. Together supports 200+ base models on the inference side. kolm targets a smaller curated set (Phi-3, Qwen2.5, Llama-3.2 in the MIT/Apache footprint) - the bases we've validated end-to-end through the verifier and distillation pipeline.

Quick iteration. Together's eval + redeploy loop is a few clicks. kolm asks you to re-compile, which takes longer per cycle (though you get a fresh file out of it).

Where kolm wins

You own the file. The deliverable is portable. Run it on your laptop, ship it to a customer, deploy it air-gapped, embed it in a phone app. None of that is possible when the model is on Together's servers.

$0 marginal inference. Per-token billing scales linearly with traffic; flat-compile + own-hardware does not. For high-volume tasks the curves cross fast.

Verifier-gated training. kolm runs every captured pair through a synthesized verifier before it lands in the training set. The result is higher-quality labels and a tighter K-score gate. Together does the supervised fit but doesn't ship a verifier-gating layer.

Receipts. Every .kolm output ships with an HMAC-SHA256 receipt chain. Together has no equivalent.

No sustained vendor cost. If Together's pricing changes, your fine-tune lives on a moving floor. If kolm's pricing changes, the .kolm files you already compiled keep running on your hardware.

Side-by-side

When to use Together

Use Together when you want a fine-tune and hosted inference and you don't want to think about either. The fastest exit for "I have pairs, I want a serving endpoint by Friday."

# Together fine-tune flow:
together.fine_tuning.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    training_file="file-abc",
    n_epochs=3)
# then call ft:llama-3.1-8b-instruct:org:abc through chat completions

When to use kolm

Use kolm when you want the model itself - signed, portable, $0-marginal, runs on your hardware. The right answer for high-volume tasks, regulated industries, on-device deployment, or any situation where vendor lock and per-call billing are friction.

# point traffic at the kolm capture proxy:
ANTHROPIC_BASE_URL=https://kolm.ai/v1/capture/anthropic

# once enough pairs accumulate, compile:
kolm compile "summarize support tickets" \
  --namespace support \
  --base qwen2.5-7b

ok wrote support.kolm  k_score=0.89  signature=hmac-sha256

Can I use both?

Yes. Together is a reasonable inference backend for the frontier-grade tier of your stack while you compile high-volume sub-tasks into .kolm files for the local tier. The capture proxy preserves the upstream call, so the Together endpoint keeps doing its job until the local artifact is ready.