VORIM
We use cookies

We use cookies to analyze site traffic and improve your experience. You can choose to accept all cookies or only essential ones. See our Privacy Policy.

Voram 1

The open model that decides, in real time, whether an AI agent may take an action. Allow, deny, or escalate, before it runs, tied to identity and a signed audit trail.

We are building it in the open. Join the waitlist for the weights, the benchmark, and early access when it lands.

No spam. One email when it is ready.
voram-1 · decision42 ms
agent payments-bot agt_9f2c
action transfer.execute $8,400 → vendor_ext
policy limit $5,000 · new payee
ESCALATEabove limit, unknown payee → route to human
signed · sha256 · agt_9f2c
What Voram 1 is
What it decides

Given an agent action, its permissions, and your policy, Voram 1 reasons allow, deny, or escalate before the action ever runs.

Why a dedicated model

Authorization sits inline in front of every agent action. Even the fastest hosted frontier variants take a few hundred milliseconds before the network round-trip, so a hosted API cannot meet a sub-100ms inline budget. That is a network-plus-inference floor, not a model-speed problem. A small model you run yourself skips the hop. Voram 1 targets under 100 milliseconds.

Open weights

Voram 1 will be released open, Apache 2.0, and built on Qwen 3, so once it lands you will be able to run it with the same local tools you already use for open models, like Ollama, vLLM, or llama.cpp, on your own hardware or in your own cloud. If you would rather not host it yourself, our managed service and dashboard will be the paid plans.

The stack

Fine-tuned from an open base (Qwen 3, Apache 2.0) and served for sub-100ms inline inference, optimized with TensorRT-LLM and FP8 on NVIDIA GPUs, and small enough to run on your own hardware or in your own cloud. Built for the runtime decision path, not the demo.

Not another chatbot

Voram 1 is not a general-purpose LLM

ChatGPT, Claude, Mistral, Cohere, and the open models like Llama and Qwen are general-purpose generative models. You give them a prompt and they write, reason, and generate. They are extraordinary, and they are the brains inside modern AI agents.

Voram 1 does none of that. It does not chat, write code, or generate text. It answers one narrow question, over and over, in the moment an agent tries to act: may this agent take this action, right now, under this policy. The only outputs are allow, deny, or escalate, tied to the agent's identity and written to a signed audit trail. It is not the agent's brain. It is the gate the brain has to pass through.

General LLMs
  • Generate text, code, and reasoning
  • Open-ended, one big model for everything
  • Answer measured in seconds, and that is fine
  • The intelligence inside the agent
Voram 1
A purpose-built decision model
  • Decides allow, deny, or escalate. Nothing else
  • Small and narrow, built for one job
  • Answer in milliseconds, because it runs inline
  • The gate the agent has to pass through

This is a proven shape. Meta, Google, and IBM all ship small, dedicated guard models that sit inline and make a fast safety call on each request, rather than routing every check through a big general model. They judge the safety of content: is this text harmful, biased, or a jailbreak? Voram 1 takes that shape to a question none of them answer: is this agent action authorized, right now, under this policy, tied to the agent's identity and a signed audit trail?

Llama Guard
Meta
Is this content harmful?
ShieldGemma
Google
Is this content harmful?
IBM
Granite Guardian
IBM
Is this content harmful?
Voram 1
Vorim AI
Is this agent action authorized?

Same shape, different question. The content guards judge text. Voram 1 decides actions, tied to the agent's identity and a signed audit trail, and returns allow, deny, or escalate.

Why open

The future is open, and closed models still matter

Open models are winning the ground where trust and control matter most. When a model gets to decide whether your agent may move money, delete data, or call an external tool, you should be able to open it up, see how it decides, run it on your own hardware or in your own cloud, and keep it running even if a vendor changes its mind. That is why Voram 1 is open, Apache 2.0. A gate you cannot inspect is not really a gate you can trust.

This is not a case against closed frontier models. ChatGPT and Claude are doing things no open model matches yet, and for the hardest generative work they are often the right call. We think the two coexist. Let the powerful closed models do the open-ended thinking, and let a small open model like Voram 1 stand at the gate, deciding what those agents are actually allowed to do, in the open, where anyone can check the rules.

Early results

Why a dedicated model, in numbers

We benchmarked a frontier model on our real agent-authorization decisions. It was accurate, and far too slow to run inline in front of every agent action.

benchmark · real decisionsn = 300
Frontier accuracy90%
Frontier latency964 ms
Inline auth budgetunder 100 ms
Voram 1 target≈30× faster, at parity

The point of Voram 1 is not to beat a frontier model on accuracy. The frontier model already scores well. The point is speed: a purpose-built small model can make the same decision fast enough to sit inline, on your own hardware or your own cloud, with the weights open.

Honest caveat

Voram 1 is in active development, not yet released. The first version mirrors our current policy engine at much lower latency; the deeper reasoning on novel cases grows as real decision data accumulates. We will publish the benchmark and the weights, and we would rather tell you what is real than oversell it.

Get Voram 1 when it lands.

Weights, benchmark, and early access. One email, no spam.