Released v1.0.0October 11, 2026. Open weights, Apache 2.0.

Avrik-Pro-12B

The default Avrik model. About 90 ms per decision on one GPU.

The Pro tier and the default model of the Avrik family: a 12B decision model that answers in about 90 ms on one GPU, or on a CPU or 16 GB card with the GGUF build. The right starting point for most workloads.

View on Hugging Face GGUF Q8_0 build

91.3%Dev-set accuracy292 of 320, 95% CI 87.6–93.9%

90.4 msPer decision on one H100p50 at 702 tokens, bf16

86.1%JevBench public tasks199 of 231, self-run, not official

2 buildsbf16 and GGUF Q8_0Each with its own calibration

What it’s for.

Send a state, like a ticket, a transaction, a policy and a case, and one or more typed questions. Get back a calibrated probability for every option, read straight from the model’s output in one forward pass.

Routing, triage and approvals
Guardrails and policy checks in agent loops
A fast first stage in front of people or larger models

Pick a build.

Same model, same API, two ways to run it. Each build has its own repo and its own calibration, fitted on that build’s outputs.

bf16

avrik-ai/Avrik-Pro-12B

Format: bf16 safetensors
Size: 23 GiB
Runs on: One GPU via vLLM
Context: 70,000 tokens
Latency: 90.4 ms p50 (94.6 ms p95) at 702 tokens on an H100
Temperatures: choice 2.52, yes/no 5.0
Calibration: cal-12b-bf16-frozen-20261011-1

GGUF Q8_0

avrik-ai/Avrik-Pro-12B-GGUF

Format: GGUF Q8_0
Size: 12 GiB
Runs on: llama.cpp on a CPU or a 16 GB card
Context: 4,096 tokens
Latency: First decision in 23.7 s on a 6-thread laptop CPU, including model load
Temperatures: choice 2.84, yes/no 5.0
Calibration: cal-12b-q8-frozen-20261011-1

On the same 128 calibration items, Q8_0 and bf16 both got 105 right, and the top answer agreed on 126.

Batching: 32 questions on one 1,500-token state took 17.6 ms per decision.

Four question types. One API.

All three Avrik models answer the same /v1/systemone API with the same question types. Pick a tier by accuracy, latency, and hardware, and switch without changing your code.

Type	You send	You get
Yes or no `noul`	A yes/no question	P(yes), from 0 to 1
Choice `choice`	A question and two or more named options	The top option and a probability for each
Score `score`	A question and ordered levels, like low, medium, high	The expected level and a probability per level
Multi-label `multi_label`	A question and a set of labels	P(yes) per label, and the labels above a threshold

Confidence on every answerEach answer carries top, margin, confidence, and an abstain flag.
Room to say “I don’t know”With abstain_option, the model can answer “Insufficient information to decide.”
ReplayableEight audit fields per response. The same request on the same backend and precision returns the same numbers.

How an answer is made.

Run the model through avrik-server, not generate(). The readout and calibration are what turn raw weights into probabilities you can act on.

1
Prompt renderer
One fixed template per build, its SHA-256 in every response. Options become letters A, B, C.
2
Gemma 4 decoder
One forward pass per question. Questions on the same state are batched. Nothing is generated.
3
Letter readout
Reads the probability of every form of each option letter, and refuses if the letters hold under 98% of the mass.
4
Calibration
Per-type temperatures fitted for this exact build and bound to its hash.

Results, with their error bars.

Measured by Avrik AI on the bf16 build. Every accuracy carries a 95% Wilson interval, because a bare point estimate hides how few items sit behind it.

Avrik dev set

292/320 overall. 320 text items in 24 families. 100 labels are computed by code; 220 were drafted by a model and aren’t human-verified, so treat those families as indicative.

Agent governance

Prompt injection15/15

Intent alignment14/15

Personal data15/15

Tool-call authorization15/15

Core reasoning

Abstains when ambiguous10/10

Extraction10/10

Judgement15/15

Long policy10/10

Multi-hop10/10

Paraphrase edits12/15

Probability laws14/15

Routing10/10

Rule precedence15/15

Safety judgement15/15

Temporal and numeric20/24

Adversarial traps8/10

Long context

Buried policy clause7/16

Long records18/18

Needle in a haystack12/12

Consistency

Near-tie stability10/10

Negation pairs15/15

Order invariance8/10

Paraphrase invariance10/10

Scoring

Score quality4/10

Accuracy per family, dot is the point estimate, line is the 95% Wilson interval.

JevBench public tasks

The 231 public tasks of JevBench v1.4.2, run by Avrik AI. Not an official measurement, and not on the scale of the current leaderboard.

Easy48/48

Standard67/72

Hard84/111

Overall199/231

Top-answer accuracy by difficulty tier, with 95% Wilson intervals. Axis starts at 50%.

Held-out human evaluation

Pending. Avrik’s sealed evaluation set is written by people, kept encrypted, and never used for training or tuning. Results will be added to the model card in a later revision.

Calibration

Expected calibration error, 10 bins, top option. Lower is better. When this model says 0.8, it should be right about 80% of the time. Calibration pool held out from dev; dev shown for comparison. The pool’s yes/no items come from two hard families, which is why the yes/no temperature sits at the top of the fit range (5.0). It is refitted when the pool widens.

Calibration pool, choice n = 1260.093 0.047

Calibration pool, yes/no n = 840.283 0.160

Dev set, choice n = 1500.013 0.022

Dev set, yes/no n = 1600.103 0.062

Dev set, score n = 100.600 0.571

bf16 build, calibration cal-12b-bf16-frozen-20261011-1.

GGUF Q8_0 build

Calibration pool, choice n = 840.121 0.087

Calibration pool, yes/no n = 440.270 0.188

GGUF Q8_0 build, calibration cal-12b-q8-frozen-20261011-1, fitted on this build’s own outputs.

Run it in a few commands.

avrik pull verifies every file before anything loads. avrik serve exposes the API on 127.0.0.1:8080, loopback only.

On a GPU with vLLM (Linux, CUDA). avrik pull downloads the repo and verifies every file, vLLM serves the weights on loopback, and avrik serve puts the decision API in front.

pip install avrik "vllm==0.30.0"
avrik pull Avrik-Pro-12B --format bf16          # -> ~/.cache/avrik/Avrik-Pro-12B-bf16

HF_HUB_OFFLINE=1 VLLM_NO_USAGE_STATS=1 DO_NOT_TRACK=1 VLLM_BATCH_INVARIANT=1 \
  vllm serve ~/.cache/avrik/Avrik-Pro-12B-bf16 --served-model-name avrik \
  --dtype bfloat16 --max-model-len 70000 --logprobs-mode raw_logprobs \
  --host 127.0.0.1 --port 8000 &

avrik ask Avrik-Pro-12B --format bf16 --vllm-url http://127.0.0.1:8000 \
  --state "Customer bought headphones 12 days ago, unopened, receipt attached. \
Policy: unopened items within 30 days get a full refund." \
  --noul "Should the refund be approved?" --choice "Which action?=approve,deny,escalate"

avrik serve Avrik-Pro-12B --format bf16 --vllm-url http://127.0.0.1:8000   # API on :8080

On a CPU (Linux, macOS or Windows; no GPU needed). The first line installs llama.cpp’s prebuilt CPU wheel, so nothing has to compile.

pip install llama-cpp-python==0.3.35 --no-deps --only-binary=:all: \
  --index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
pip install "avrik[cpu]"

avrik pull Avrik-Pro-12B                    # downloads avrik-ai/Avrik-Pro-12B-GGUF and verifies it
avrik ask Avrik-Pro-12B \
  --state "Customer bought headphones 12 days ago, unopened, receipt attached. \
Policy: unopened items within 30 days get a full refund." \
  --noul "Should the refund be approved?" \
  --choice "Which action?=approve,deny,escalate"
avrik serve Avrik-Pro-12B                   # the API on http://127.0.0.1:8080

Route	Purpose
`POST /v1/systemone`	Decide: state and questions in, calibrated answers out
`GET /health`	Readiness and the loaded build’s hashes
`GET /v1/models`	The served model
`GET /openapi.json`	The OpenAPI 3.1 schema

# Typed Python client, sync and async
pip install avrik-sdk

Use it well.

Set your own thresholds: approve above 0.95, send to a person between 0.2 and 0.95, and so on.

Built for

Automated decisions with a threshold a person sets
Triage, routing, policy and eligibility checks
Classification with calibrated confidence
A fast first stage in front of people or larger models

Not built for

Explanations or free text. It generates none.
Acting alone on legal, medical, financial, or safety decisions without human review
Images, in this release. Input is text or JSON.
Languages other than English, which haven’t been evaluated

Weakest dev-set areas: score_quality (4/10), long_policy_buried (7/16), order_invariance (8/10). The model judges only what’s in the state, so instructions planted there can sway it. Keep untrusted text clearly delimited.

Provenance you can check.

Every file is listed with its SHA-256 in MANIFEST.sha256. Inference makes no network calls, sends no telemetry, and stores nothing.

Base	`google/gemma-4-12B-it` at `707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7`
Recipe	Frozen base with fitted per-type temperatures. No fine-tuning.
Release gate	Frozen base. Fine-tuned builds replace this one only if they beat it on the sealed, human-authored set.
bf16 build	Model hash `avrik-ai/Avrik-Pro-12B@cfd9cc0c27e96d1e`Prompt template `8d1575803fdb25b66865405afa0d2e6bb3df4eff08003e858769225b3bcc0b70`
GGUF Q8_0 build	Model hash `avrik-ai/Avrik-Pro-12B-GGUF@6fbf37d64038ea71`Prompt template `3269884bf2b10959761578efcec48f09ec59ac185c735f0f759a9088440e2ce0`
Release	`v1.0.0`, October 11, 2026

Tamper checksSwapped or altered downloads are refused, and the manifest’s own digest is pinned in the avrik package.
No code in releasesCode files, auto_map, and symlinks are refused. Tokenizers load with trust_remote_code=False.
LicenceApache 2.0 for the weights, calibration, and readout config. Keep NOTICE with any redistribution.

The rest of the family.

Same API, same question types. Move up a tier for accuracy, or down for speed and cost.

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
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

Get Avrik-Pro-12B on Hugging Face.

Open weights under Apache 2.0, with calibration and hashes for every build. Questions or a use case to talk through? We’d like to hear it.

View on Hugging Face All Avrik models

Avrik-Pro-12B

What it’s for.

Pick a build.

Four question types. One API.

How an answer is made.

Prompt renderer

Gemma 4 decoder

Letter readout

Calibration

Results, with their error bars.

Avrik dev set

JevBench public tasks

Held-out human evaluation

Calibration

GGUF Q8_0 build

Run it in a few commands.

Use it well.

Built for

Not built for

Provenance you can check.

The rest of the family.

Avrik-Lite-4B

Avrik-Max-31B

Get Avrik-Pro-12B on Hugging Face.