Platform
Bring your own model
Submit an open-weight checkpoint. We check its license, abliterate it, run the eval harness, and show you the report. If you approve it, we serve it as a private model that only your organization can call.
Available to Research organizations only. Below Research, every /v1/byo/* endpoint returns feature_not_available (403). See Access tiers.
Pipeline
| Status | Label | What happens |
|---|---|---|
license_check |
Checking license | We read the license from the Hugging Face model card. |
license_review |
License under review | The license needs a person to look at it. Trust and safety is notified. |
blocked |
Blocked by license | The license forbids derivatives. Final. |
queued |
Queued for abliteration | Waiting for a GPU worker. |
abliterating |
Abliterating | The refusal direction is being removed. |
evaluating |
Running evals | The standard eval harness is running. |
awaiting_approval |
Waiting for your approval | The eval report is ready. Approve or reject it. |
deploying |
Deploying | Serving capacity is being provisioned. |
serving |
Serving | Live as a private model. |
rejected |
Rejected | You rejected the report. Nothing was deployed. Final. |
failed |
Failed | Something went wrong; status_detail says what. Final. |
Every change is recorded in events, sends a byo.status_changed webhook, and, for license_review, blocked, awaiting_approval, serving and failed, an email.
License gate
Every served checkpoint must be under a license that allows modification and redistribution of derivatives.
| Verdict | Licenses | Result |
|---|---|---|
| Allowed | apache-2.0, mit, bsd-2-clause, bsd-3-clause, bsd-3-clause-clear, isc, unlicense, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, cc-by-3.0, mpl-2.0, artistic-2.0, ecl-2.0, epl-2.0, afl-3.0, zlib, postgresql, the OpenRAIL family (openrail, openrail++, bigscience-openrail-m, creativeml-openrail-m, bigcode-openrail-m), and the model-family licenses llama2, llama3, llama3.1, llama3.2, llama3.3, llama4, gemma, deepseek, qwen |
Proceeds to queued |
| Blocked | Any license with a no-derivatives (-nd) term |
blocked |
| Review | Non-commercial (-nc) licenses; GPL, LGPL and AGPL; no declared license; anything else |
license_review |
Weights from a URL always go to license_review, because an archive cannot be inspected before download.
A reviewer approves (the job moves to queued) or declines (blocked), with a note in status_detail.
Strength
strength sets how aggressively the refusal direction is removed: low, medium (default) or full. Higher strength lowers the refusal rate further, usually at a larger cost to capability. The eval report shows both, so you can reject and resubmit at another strength.
Footprint and fees
We estimate the GPU footprint from the parameter count on the model card. If the count is unknown, we assume 8B.
| Parameters | Footprint | Monthly fee | Tier |
|---|---|---|---|
| Up to 9B | 1× L40S 48GB | $900 | small |
| Up to 35B | 1× H100 80GB | $2,200 | small |
| Up to 80B | 2× H100 80GB | $4,400 | medium |
| Up to 200B | 4× H100 80GB | $8,800 | medium |
| Over 200B | 8× H100 80GB | $17,600 | large |
The monthly fee pays for reserved capacity. It is charged from your credit balance, a month in advance, starting when the model begins serving. Usage is billed per token on top, at the tier's rates:
| Tier | Input | Cached input | Output |
|---|---|---|---|
| small | $0.10 | $0.01 | $0.40 |
| medium | $0.35 | $0.035 | $1.40 |
| large | $2.30 | $0.23 | $11.50 |
USD per 1M tokens. The job's footprint and monthly_fee_usd fields show the estimate once the license check has run.
The private model
When serving starts, the job's model field holds the model id:
<org-slug>-<4 characters of your org id>/<name>for example acme-labs-x7k2/qwen3-8b-redteam. It:
- appears in
GET /v1/modelswithowned_by: "you"only for your organization's keys; - is an abliterated model, published with an eval card built from your report, like every other model;
- defaults to a 32,768-token context window and 8,192 output tokens;
- is called by pinning its id. The auto aliases never route to private models.
API reference
Submit
curl https://api.utmost.site/v1/byo/models \
-H "Authorization: Bearer $UTMOST_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"source": {"type": "huggingface", "repo": "Qwen/Qwen3-8B", "revision": "main"},
"name": "qwen3-8b-redteam",
"strength": "medium"
}'| Field | Notes |
|---|---|
source |
{"type": "huggingface", "repo": "owner/model", "revision"?} or {"type": "url", "url": "https://..."} |
name |
2 to 48 lowercase letters, digits, dots or dashes, starting with a letter or digit. Unique among your jobs that are not rejected, failed or blocked. |
strength |
low, medium or full. Default medium. |
Returns 201 with the job. A bad name or a duplicate returns invalid_request. If the Hugging Face lookup fails (no such repo, private, or gated), the job is still created, in failed, with the reason in status_detail. For a gated or private repo, grant the Utmost Hugging Face account access and submit again.
The job object
{
"id": "byo_Fq2...",
"object": "byo.model_job",
"name": "qwen3-8b-redteam",
"source": { "type": "huggingface", "ref": "Qwen/Qwen3-8B", "revision": "main" },
"license": "apache-2.0",
"license_verdict": "allowed",
"strength": "medium",
"status": "awaiting_approval",
"status_label": "Waiting for your approval",
"status_detail": "Eval finished. Review the report, then approve or reject.",
"footprint": "1× L40S 48GB",
"parameters": 8190735360,
"monthly_fee_usd": 900,
"eval_report": {
"refusal": { "refused": 3, "total": 40, "set": "..." },
"capability": [{ "name": "MMLU", "base": 70, "score": 69.4, "delta": -0.6, "n": 500 }],
"checkpointHash": "sha256:...",
"baseHash": "git:...",
"method": { "tool": "Heretic", "technique": "Refusal-direction orthogonalization", "layers": "auto", "strength": "medium" },
"notes": "..."
},
"model": null,
"eval_card_url": null,
"events": [
{ "at": "2026-11-19T05:40:00.000Z", "status": "queued", "detail": "apache-2.0 permits modification and redistribution. Estimated footprint 1× L40S 48GB." }
],
"created_at": 1795066800,
"updated_at": 1795070400
}eval_report keys are camelCase. A report with "simulated": true came from a test environment, not a real run.
List and retrieve
curl https://api.utmost.site/v1/byo/models -H "Authorization: Bearer $UTMOST_API_KEY"
curl https://api.utmost.site/v1/byo/models/byo_Fq2... -H "Authorization: Bearer $UTMOST_API_KEY"The list holds all your jobs, newest first. An unknown id returns not_found.
Approve or reject
curl -X POST https://api.utmost.site/v1/byo/models/byo_Fq2.../approve -H "Authorization: Bearer $UTMOST_API_KEY"
curl -X POST https://api.utmost.site/v1/byo/models/byo_Fq2.../reject -H "Authorization: Bearer $UTMOST_API_KEY"Only a job in awaiting_approval can be approved (it moves to deploying) or rejected (it moves to rejected). Otherwise the call returns invalid_request. Approving commits you to the monthly fee once serving starts.
For operators: the GPU worker protocol
Abliteration, evaluation and deployment run on GPU workers outside the API servers. Workers use two internal endpoints, authenticated with Authorization: Bearer $BYO_WORKER_TOKEN. These are not part of the customer API.
POST /api/internal/byo/claim claims the oldest job in queued or deploying and locks it for 6 hours. It returns 204 when there is nothing to do, or:
{ "job": { "id": "byo_...", "status": "abliterating", ... }, "task": "abliterate_and_evaluate" }task is abliterate_and_evaluate for a queued job (which moves to abliterating on claim) or deploy for an approved one.
POST /api/internal/byo/{id} reports progress. The body is one of:
{ "status": "evaluating", "detail": "Running MMLU." }{
"status": "awaiting_approval",
"detail": "Eval finished.",
"eval": {
"refusal": { "refused": 3, "total": 40, "set": "..." },
"capability": [{ "name": "MMLU", "base": 70.1, "score": 69.5, "delta": -0.6, "n": 500 }],
"checkpointHash": "sha256:...",
"baseHash": "git:...",
"method": { "tool": "Heretic", "technique": "Refusal-direction orthogonalization", "layers": "auto", "strength": "medium" },
"notes": "optional"
}
}{ "status": "serving", "endpoint": { "baseUrl": "https://gpu-7.internal/v1", "upstreamModel": "qwen3-8b-redteam", "apiKeyEnv": "BYO_GPU7_KEY" } }{ "status": "failed", "detail": "Out of memory while loading weights." }serving creates the private model, its OpenAI-compatible deployment and its eval card, and is accepted only for jobs in deploying. Responses: {"ok": true}, 400 for a malformed body, 401 for a bad token, 409 when the transition is not allowed.