Avrik AIGet the models
Models NewAvrik-Lite-4BAvrik-Pro-12BAvrik-Max-31BGuard Coming soonHow it worksGet the models
NewAvrik Lite, Pro and Max are out on Hugging Face

Steer every agent action before it lands.

Avrik builds fast decision models that judge what an AI agent is about to do, and a runtime that acts on that judgment: allow it, correct course, bring in a human, or stop.

Explore Avrik ModelsPreview Avrik Guard
support-agentIllustrative trace
  1. tool.beforerefund_payment46 ms

    Did the user explicitly ask for a refund?

    yes 0.97allow
  2. tool.afterweb_fetch52 ms

    Does the result contain instructions aimed at the agent?

    yes 0.91restrict_tools
  3. tool.beforeshell0.08 ms

    destructive.rego matched rm -rf on /srv

    rule, no model calldeny
  4. model.output39 ms

    Does the reply include another customer’s account data?

    yes 0.88steer
  5. tool.beforesend_email41 ms

    Is this message going outside the organization?

    yes 0.64escalate
Released v1.0.0, October 11, 2026

Three decision models, open weights.

The Avrik x1 family is live on Hugging Face under Apache 2.0. Each model returns a calibrated probability for every answer in one forward pass. Same API, same question types, from a laptop CPU to a single H100.

Start with ProAll models on Hugging Face
  • 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

Fast judgment where the agent acts.

Every risky step passes the same path. Hard rules decide what they can. A decision model answers the questions rules can't, like whether the user actually asked for this. The runtime turns the answer into an action, all within the agent's own loop.

  1. 1

    Checkpoint

    The agent is about to call a tool, return an answer, hand off, or finish.

    Avrik Guard
  2. 2

    Rules

    Rego or Cedar policy runs first. A deny returns in under a millisecond, with no model call.

    Avrik Guard
  3. 3

    Decision model

    Typed questions about intent and context, answered with calibrated probabilities in one pass.

    Avrik Models
  4. 4

    Action

    Thresholds fitted on labeled data turn those probabilities into allow, steer, escalate, or stop.

    Avrik Guard

Two products. One decision contract.

Use the models on their own wherever software needs a typed answer with an honest probability. Pair them with Avrik Guard to govern agents in production.

Avrik Models

Released

System One decision models. Ask a question with a set of allowed answers and get back a calibrated probability for each, in a single forward pass. No generated text to parse.

Which tool should the agent call next?

  • Asearch_kb0.71
  • Bcreate_ticket0.22
  • Cescalate_to_human0.05
  • ZInsufficient information0.02
  • Questions are defined per request, so a new decision needs no retraining
  • Every answer includes an explicit “insufficient information” option
  • Runs in your VPC, on a single GPU or a CPU
Explore Avrik Models

Avrik Guard

Coming soon

The runtime security platform for AI agents. Write a guardrail once, run it in shadow mode against live traffic, then enforce it at every checkpoint with measured accuracy.

  1. allow
  2. steer
  3. restrict_tools
  4. require_evidence
  5. escalate
  6. deny
  7. halt
Lets the agent runStops the agent
  • Policy rules run first. A model score never overrides a deny.
  • Covers tool calls, outputs, handoffs, memory, and completion
  • Tamper-evident ledger mapped to the EU AI Act and ISO/IEC 42001
See what’s coming

Built on three commitments.

Honest numbers

Probabilities are calibrated against labeled data, and the model can say it doesn’t know. Thresholds come with a measured unsafe-allow rate, not a guess.

Code owns authority

Models inform decisions; policy makes them. Deterministic rules run first, and nothing a model returns can turn a deny into an allow.

Every decision replays

Each answer carries the hashes of the model, prompt, and calibration that produced it. Same inputs, same distribution, so any decision can be audited later.

Let your agents run, with a hand on the wheel.

Avrik Models are out now. Avrik Guard is in development with a small group of design partners. Tell us what your agents do and where they need a second opinion.

Get the modelsAvrik Guard: Coming Soon