bev: open AI decision model by Berget AI

Q: How do I run bev?

Pull the weights with systemone pull berget-ai/bev and load them with the runtime named in the manifest.

bev-1

Berget AI's System One model for gating agent commands: a LoRA adapter and a fine-tuned joint schema head on Cloudflare's Clef-Flash that answer noul, choice and score questions over a state in one forward pass. Trained on Swedish and English operations decisions.

A rank-16 LoRA (about 116 MB) on Cloudflare's Clef-Flash plus a replacement joint head; it needs the base weights and Cloudflare's joint_schema_model.py. Trained on 221,759 judged operations decisions (risk and privacy, memory, routing, evasion), with labels checked by a larger model and human review. Berget's own held-out results come from the training corpora: risk 97.0% (base 93.5%), evasion 93.3% on 90 cases, red team 74% on 34 commands, and 82.8% on a general Jev-style set of 1,200 (base 80.6%); no calibration claim. It is the model behind Berget's MIT-licensed guardrails-md gate for coding agents.

What it decides

noul — answers a yes/no question with one probability
choice — picks one option from a set
score — places the input on an ordered scale

At a glance


Parameters	9B
Base model	`Cloudflare/clef-flash`
Maker	Berget AI
Released	2026-10-04
License	apache-2.0
Reported accuracy	97.0%

Get the weights

pip install systemonemodels
systemone pull berget-ai/bev

The files are served from the maker's Hugging Face repository, bergetai/bev-1, and verified against the checksums recorded here.

Playground

Build a request for berget-ai/bev. The form and the JSON are the same object; the response panel shows the shape every System One server returns.

Input

StateThe context the questions are asked about. Up to the model's context length.

A yes/no question. The answer is one probability that the condition holds.

Instructions

Response

Placeholder values, in the shape a run returns.

safenoul7%

This model doesn't answer live here. Send this request to your own server for this model; every System One server speaks this shape.

Call it

curl https://api.systemonemodels.ai/v1/systemone \
  -H "Authorization: Bearer $SYSTEMONE_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"berget-ai/bev","state":"Task: clean up inactive accounts before the quarterly report.\nProposed tool call: delete_rows(table=\"customers\", where=\"last_login < 2023-01-01\")\nContext: the customers table is production; there is no backup newer than Friday.","questions":{"safe":{"type":"noul","instructions":"Is this action safe to run without a human approving it first?"}}}'

Providers

Where this model runs today. Call the provider directly or run the weights yourself.

Provider	Status	Input	Output	Context	Link
Your own hardware	Open weights	Free	Free	—	`systemone pull berget-ai/bev`

Pricing

Availability: Open weights
Licence: apache-2.0
Hosted API: None recorded. Pull the weights and run it yourself.

Pulling and running the weights costs nothing beyond your own hardware; the licence above says what you may do with them.

Performance

Decision accuracy	97.0%

Suite: Berget held-out risk split (16,902 questions; same operations-traffic corpora as training). Numbers are the publisher's own, from the manifest; the model card says how they were measured.

adapter_config.json1.1 kB
adapter_model.safetensors9236b43111 MB
joint_head.safetensorsba28c88228 MB

Every file resolves through a stable path — /v1/models/berget-ai/bev/resolve/2026.09/<path> — while the object underneath is content-addressed and shared with every other version that holds the same bytes. API reference.

Model tree

Every ancestor down to this model, then what was built on it, wherever each one lives.

Activity

Downloads0

Stars0

Versions1

Built with it0

Latest version: 2026.09
Updated: Oct 7, 2026
Added to the registry: Oct 7, 2026
Mirrors: bergetai/bev-1

Documentation

The HTTP API — every endpoint the website and CLI use.
The CLI — pull this model with systemone pull berget-ai/bev.
systemone.yaml — what the manifest below records.

systemone.yaml

{
  "spec_version": "0.1",
  "model": "bev",
  "namespace": "berget-ai",
  "category": "system-one",
  "architecture": "clef",
  "base_model": "cloudflare/clef-flash",
  "base_model_source": "huggingface",
  "license": "apache-2.0",
  "license_name": null,
  "summary": "Berget AI's System One model for gating agent commands: a LoRA adapter and a fine-tuned joint schema head on Cloudflare's Clef-Flash that answer noul, choice and score questions over a state in one forward pass. Trained on Swedish and English operations decisions.",
  "tags": [
    "system-one",
    "lora",
    "clef",
    "guardrails",
    "safety",
    "swedish",
    "9b"
  ],
  "capabilities": [
    "noul",
    "choice",
    "score"
  ],
  "runtime": {
    "framework": "pytorch",
    "entrypoint": null,
    "python": null,
    "hardware": "gpu"
  },
  "artifacts": [],
  "evaluation": {
    "suite": "Berget held-out risk split (16,902 questions; same operations-traffic corpora as training)",
    "suite_version": null,
    "decision_accuracy": 0.97,
    "calibration_error": null,
    "valid_action_rate": null,
    "median_latency_ms": null,
    "p95_latency_ms": null
  },
  "parameters": 9000000000,
  "context_length": null,
  "hub_repo": "bergetai/bev-1",
  "released": "2026-10-04",
  "availability": "open-weights",
  "hosted": null
}

View raw metadata

FAQ

What does bev decide?

It answers 3 kinds of question: noul returns the probability that a condition holds; choice picks one option from a list you give it; score rates the state on a scale you define. Every answer is a typed value with a probability, never free text.

How do I run bev?

Pull the weights with systemone pull berget-ai/bev and load them with the runtime named in the manifest.

Can I use bev commercially?

It is published under apache-2.0. Read the licence before shipping; the model card may add terms of its own.

What is bev built on?

Its base model is cloudflare/clef-flash, a clef architecture. The Model tree section shows the whole chain.

How much can bev read at once?

The publisher has not recorded a context length. The model has 9.0B parameters.

How accurate is bev?

Its publisher reports 97.0% decision accuracy on Berget held-out risk split (16,902 questions; same operations-traffic corpora as training). Suites differ between publishers, so compare models on your own labelled examples.

How do I link to bev?

This page is https://systemonemodels.ai/berget-ai/bev. Its metadata is at https://api.systemonemodels.ai/v1/models/berget-ai/bev and a Markdown summary for agents at https://systemonemodels.ai/berget-ai/bev.md.

Berget AI: bev

bev-1

What it decides

At a glance

Get the weights

Read more

Playground

Providers

Pricing

Performance

Model tree

Activity

Documentation

systemone.yaml

FAQ

Berget AI: bev

bev-1

What it decides

At a glance

Get the weights

Read more

Playground

Providers

Pricing

Performance

Model tree

Activity

Documentation

systemone.yaml

FAQ