Transparent abliteration, for authorized research

Refusal-free models.Published proof.

One API for abliterated open models. Every checkpoint ships with an eval card that shows what was removed, and what it cost.

Base and abliterated pairs of

01The reflex

Standard models refuse the work. Red teams can’t probe. Moderation teams can’t label. Security researchers can’t review their own code. The work is authorized. The refusal is a reflex.

02The method

A refusal is a direction.

Inside the model, the urge to refuse runs along one measurable axis. Call it r̂.

We project it out.

Refusal-direction orthogonalization, run with an open-source engine. Layer by layer. Nothing retrained.

Then we measure what’s left.

Reasoning, maths, code, long context. All scored against the base model before a single request is served.

03The proof

Every checkpoint ships with its receipt.

An eval card is published before a model is served. Not after, and not on request.

  • Refusal rate on a stated prompt set
  • Capability kept, as deltas against the base
  • Where the model got worse
  • Tool, technique, layers, strength, and both hashes
Eval card · No. 014Specimen

utmost/kimi-k3.1-pro-abl

base · kimi-k3.1-pro

0/40refusals
harmful-prompt set, 40 prompts

Capability kept, Δ vs base

MMLU−0.4
GSM8K−0.7
Coding−1.2
Long context−0.2
Got worse
Wordier on multi-step maths
Method
Orthogonalization · layers 14–38 · full
Checkpoint
sha256:41bd…07c3

Specimen card. Figures are illustrative until launch.

04The route

Easy tasks never touch the large tier.

Call auto. A small classifier scores how hard the task is, and whether it needs a refusal-free model at all. You pay for the model that ran.

POST /v1/chat/completions

{
  "model": "auto",
  "messages": [{
    "role": "user",
    "content": "Summarise this changelog for the release notes."
  }]
}
reqclassifierSmall7–30BMediumFlash-classLargeKimi-class
x-utmost-route
auto → small
variant
standard

Nothing here to refuse. It goes to the standard variant, which is cheaper and scores higher.

05Your weights

Bring your own model.

  1. 01Link or uploadA Hugging Face link, or the weights themselves.
  2. 02License checkBlocked if the terms forbid modification.
  3. 03Pick a strengthLow, medium or full.
  4. 04Abliterate, then evaluateOpen-source engine. Our standard harness.
  5. 05Review and approveYou read the eval report before anything is served.
  6. 06ServeLive as tenant/model-name, private to your org.

06The limits

Two limits no tier removes.

  • Child sexual abuse material. Blocked, and reported as the law requires.
  • Weapons of mass harm. Serious uplift for biological, chemical, nuclear or radiological weapons is blocked.

Both checks run in memory at the gateway. Prompts and outputs are not stored.

Access is verified, in three tiers

DeveloperVerified email and payment methodStandard models and the router
VerifiedOrg identity and a signed acceptable-use policyAbliterated checkpoints
ResearchKYB, a named use case, a contractYour own models, custom retention

07The price

Priced per token. Saved per task.

No subscription. The savings come from routing and caching, not from a signup bonus.

Abliterated checkpointsDirect price + 10–20%
Standard modelsAt or near direct price
Cached inputAbout 10% of input, on every model
Batch50% off, within 24 hours
The auto routerNo surcharge
CreditsPrepaid. Roll over for 12 months

Do the work.

Sign in with Google to get a key. Refusal-free checkpoints unlock once your organization is verified.