Released v1.0.0, October 11, 2026

A typed answer with an honest probability.

Avrik Models are System One decision models. Give one some context and a question with its allowed answers. It returns a calibrated probability for every answer in a single forward pass. No text is generated, so there’s nothing to parse and nothing to trust blindly.

Choose a model View on Hugging Face

90.4 msMedian latency for 700 tokens on one H100Avrik-Pro-12B, bf16 build

1 passOne forward pass per question, with no tokens generated

$0.02–0.06Per 1,000 decisions when self-hostedDesign estimate

Apache 2.0Open code, weights, and adapters

The Avrik SystemOne Model family.

Three tiers built on Google’s Gemma 4, each calibrated for every build. All three answer the same API with the same question types, so pick by accuracy, latency, and hardware.

Edge4B
Avrik-Lite-4B
Calibrated decisions on an ordinary CPU.
Dev-set accuracy
86.6%
Per decision on a laptop CPU
5.0 s
- GGUF Q4_K_M
- bf16
Hugging Face
Pro, the default12B
Avrik-Pro-12B
The default Avrik model. About 90 ms per decision on one GPU.
Dev-set accuracy
91.3%
Per decision on one H100
90.4 ms
- bf16
- GGUF Q8_0
Hugging Face
Max31B
Avrik-Max-31B
The most accurate Avrik model, for the hardest calls.
Dev-set accuracy
96.9%
Per decision on one H100
168.4 ms
- FP8
Hugging Face

Model	Base	Builds	Dev-set accuracy	JevBench public	Latency
Avrik-Lite-4BEdge	`gemma-4-E4B-it`	GGUF Q4_K_M bf16	86.6%95% CI 82.4–89.9%	78.8%	5.0 sp50 at 256 tokens, GGUF Q4_K_M
Avrik-Pro-12BPro	`gemma-4-12B-it`	bf16 GGUF Q8_0	91.3%95% CI 87.6–93.9%	86.1%	90.4 msp50 at 702 tokens, bf16
Avrik-Max-31BMax	`gemma-4-31B-it`	FP8	96.9%95% CI 94.3–98.3%	91.8%	168.4 msp50 at 702 tokens, FP8

Measured by Avrik AI. Dev set: 320 items in 24 families. JevBench: the 231 public tasks of v1.4.2, self-run, not an official measurement. Details and intervals are on each model’s page.

Classifiers can’t learn new questions. LLMs answer in prose.

A decision model is a classifier whose classes are set at request time. You get the flexibility of a language model with the shape of a classifier: a probability for each allowed answer.

Property	Trained classifier	LLM as judge	Avrik decision model
A new question	Needs new labels and retraining	Needs a new prompt	Needs a new prompt
What you get back	Fixed labels	Prose you have to parse	A probability for every allowed answer
Confidence	Often uncalibrated	Self-reported, if at all	Calibrated per answer type, with an explicit abstain
Speed	Fast	Seconds of generation	Tens of milliseconds
Audit	Varies	Hard to reproduce	Replayable from eight audit hashes

Typed questions cover most decisions.

Yes or no, choice, and score, plus multi-label for several independent yes/no reads at once. Every question can include a final option: “Insufficient information to decide.” When the evidence isn’t there, the model says so instead of guessing.

noul

Yes or no, with a probability

Did the user explicitly ask for a refund?

yes0.94

choice

One of several options

Which team should own this ticket?

billing0.78
shipping0.15
support0.05

score

A point on an ordered scale

How urgent is this, from 1 to 5?

10.02
20.08
30.31
40.46
50.13

One request. Every answer, scored and signed.

The server speaks the /v1/systemone wire protocol, so existing System One clients can switch with a URL change. Each response carries the hashes needed to replay it exactly.

Request

POST /v1/systemone
{
  "model": "Avrik-Pro-12B",
  "state": {
    "message": "I was charged twice for order 4417.
                Please refund the duplicate.",
    "next_tool": "refund_payment"
  },
  "questions": {
    "asked_for_refund": {
      "type": "noul",
      "instructions": "Did the user explicitly
                       ask for a refund?",
      "abstain_option": true
    }
  }
}

Response

{
  "answers": {
    "asked_for_refund": {
      "type": "noul",
      "noul": 0.94,
      "top": 0.94,
      "margin": 0.88,
      "abstain": false,
      "avrik_confidence": 0.92
    }
  },
  "audit": {
    "model_hash": "avrik-ai/Avrik-Pro-12B@cfd9cc0c…",
    "calibration_version": "cal-12b-bf16-frozen-20261011-1",
    "prompt_template_sha256": ["8d157580…"],
    "backend": "vllm",
    "precision": "bf16"
    … 3 more audit fields
  }
}

Enterprise by default.

Built to run inside your boundary and to explain itself to your auditors.

No egress, no telemetrySelf-host in your VPC. Nothing leaves your network.
Hash-verified weightsThe server checks every weight file and never switches models silently.
Refuses rather than guessesThe readout refuses an answer when the option letters hold under 98% of the probability mass.
Calibrate on your own labelsRefit the temperature for each answer type on data from your domain.
Standard interfacesAn OpenAPI 3.1 schema at /openapi.json, and a health route that reports the loaded build’s hashes.
Typed Python SDKpip install avrik-sdk: sync and async clients that depend only on httpx and pydantic.

Put a decision model in your loop.

All three models are on Hugging Face under Apache 2.0. Start with Pro, or tell us which questions your software needs answered.

Start with Avrik-Pro-12B See Avrik Guard