Von

Von is a 395M-parameter attempt to clone the idea behind Jev and ship it as open weights you run yourself. The pitch is the same one TypeSafe made for System One models: stop spending a frontier LLM on decisions that do not need generated text. Instead of decoding tokens, Von takes a state and a set of typed questions and returns a calibrated probability distribution in a single forward pass.

The first-party repository was created on 2026-09-18 under Apache-2.0 and written in Python 3.12+. On 2026-09-20 GitHub listed 63 stars and 10 forks, with a push the same day. The weights are on Hugging Face as wfzyx/von-1.0, a safetensors repo built on a ModernBERT backbone that had 16 likes and about 1,223 downloads at the time of writing. Sixty-three stars is a very early project, not a track record.

Key Features

• Three decision primitives: Choice returns a probability distribution over mutually exclusive options, Noul returns a calibrated binary probability for a condition, and Score returns an ordinal rating across ordered levels.
• No autoregressive decoding: one forward pass, no key-value cache growth, no schema parse errors from generated JSON.
• Calibrated by training, not by prompt: Von is post-trained with a composite objective that blends cross-entropy with Brier score, then temperature-scaled, so the reported confidence is meant to track accuracy rather than look confident.
• Hardware-agnostic: the project lists CUDA, AMD ROCm on Linux, Apple Metal Performance Shaders, and multithreaded CPU paths.
• Protocol parity: von serve exposes a /v1/systemone HTTP endpoint that the project says matches the TypeSafe wire specification, so a client written for one can talk to the other.

Benchmarks

Von's numbers come from the author running the community jabr/classifier-benchmark suite, 8 tasks and 78 cases:

Read the table as a vendor figure. It is a small suite, the author wrote both the model and the comparison, and Jev still leads on raw accuracy. What Von changes is the deployment shape: no API, no per-call price, no data leaving the machine.

Use Cases

• Agent routing: decide which tool, queue, or model a request belongs to before you pay for a generative call.
• Guardrails and verification: run a binary check on model output as a cheap second opinion.
• Triage scoring: rate severity, urgency, or sentiment on an ordinal scale without a prompt template.
• Offline stacks: local-first applications that cannot depend on a hosted decision API.

Limitation: Von cannot write prose, code, or tool-call strings. It is a decision function inside software, not a chat model, and a 395M non-autoregressive classifier will not carry the breadth of a fine-tuned frontier model on messy, unseen taxonomies.

Pricing

Getting Started

pip install von-sdk          # or: uv add von-sdk

import von

result = von.decide(
    state="Database replication lag on cluster us-west-2 exceeded 45 seconds.",
    choices={
        "infrastructure": "Database, hardware, network, or server failures",
        "billing": "Invoices, payments, refunds, subscription queries",
        "feature_request": "Requests for new platform capabilities",
    },
)

For a service, von serve --host 0.0.0.0 --port 8000 starts an HTTP endpoint, and the README shows a curl against /v1/systemone with "model": "von-1.0.0". First-party entry point: the repository README and the Hugging Face model card.

Frequently Asked Questions

Is Von affiliated with TypeSafe?

No. It is an independent project that describes itself as a drop-in alternative to Jev and claims compatibility with the same wire protocol.

Can it match Jev's accuracy?

On the author's suite it trails by roughly four points of macro accuracy. It compensates with local execution and zero per-call cost.

Do I need a GPU?

No. CPU inference is supported, but the author measures about 480 ms there against about 18 ms on Apple MPS.

Alternatives

• Jev: the commercial System One model this one imitates, ahead on accuracy and behind on openness.
• LangGraph: a workflow host that calls an ordinary LLM for routing decisions.
• GLiNER2: open-weight span and classification models for extraction tasks.

Tips

1. Keep taxonomies under the primitives' comfortable cardinality and split larger ones into a two-stage route.
2. Treat confidence as a gate: automate the high-confidence head and escalate the tail to review.
3. Verify the /v1/systemone compatibility yourself before swapping it into an existing System One client.

Conclusion

Von is the open-weight answer to a question Jev raised in September 2026: do classification and routing decisions really need a generative model? A 395M non-autoregressive classifier that runs locally in about 18 ms is a reasonable no. Try it in an agent loop where you currently pay a frontier model to pick between five labels.