Run models in your environment. Decide what needs review.

What Helix is

Helix is an inference runtime: the software that loads an AI model and answers requests with it. It is not a model and not a hosted AI service. You choose an open model, you run Helix where you need it, and your applications call it over a familiar API. Alongside each response, Helix provides a confidence score your application can use to decide what happens next.

Helix is sold on its own. It is also the runtime used with Compile-Time Inference, but you do not need one to use the other.

Where it runs

Helix is licensed as a container image that you deploy. Supported targets:

• Your Kubernetes or OpenShift cluster.
• Your own cloud account (AWS, Azure or Google Cloud).
• A virtual machine, on-premises or in a private cloud.
• An air-gapped network with no outbound connection.

In production, requests go from your application to your Helix deployment. There is no Verificate-hosted service in that path. The public demo is different: it is hosted by Verificate so you can try the engine, and should only be used with non-sensitive input.

Models and hardware

Helix serves open models, including standard and mixture-of-experts designs, on CPU-only and GPU machines. Serving is tuned for each model family. Our lead production models today are GPT-OSS 120B for text and Qwen3-VL 4B for vision; other open families such as Llama, Granite, Gemma and Qwen are served by the same engine.

Which model, precision and hardware suit your workload is part of a deployment evaluation. Confirm support for your specific model and hardware with us before you plan around it. Audio input is on the roadmap and is not available today.

The API

Helix exposes an OpenAI-compatible API, so existing OpenAI client libraries can be pointed at your deployment by changing the base address.

curl https://<your-helix-endpoint>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-120b",
    "messages": [{"role": "user", "content": "Extract as JSON..."}],
    "response_format": {"type": "json_object"},
    "stream": true
  }'

The exact response fields, including how the confidence score is returned for your version, are specified in the documentation supplied with your deployment.

Using confidence signals

The confidence score is a number from 0 to 100 that estimates how likely a response is to be reliable. Helix reports the score; it does not silently rewrite answers. Your application decides what to do with it. A common pattern:

Choosing thresholds

• Collect a representative sample of real requests with known-good answers.
• Record the score for each, then look at accuracy within score bands.
• Set thresholds from the error rate you can tolerate and the review capacity you have.
• Track how often each action is taken, and re-check after any model or version change.

Repeatable runs

Shared, batched serving can produce small differences between runs, because a result can depend on what else is being processed at the same time. For audit and compliance workloads, Helix offers a repeatable serving path — one request at a time with a fixed seed, or CPU serving — intended to return the same output for the same input. Throughput is lower in this mode. Verify repeatability on your own deployment and model before depending on it.

Performance evidence

Published measurements state their hardware, model, harness and date. They describe those configurations, not yours:

• Serving latency measured with the vLLM project’s GuideLLM harness on AMD EPYC 9254 — see Benchmarks.
• GPT-OSS 120B throughput on CPU-only AMD Zen5 EPYC systems, run by AMD engineering — see the product page.

Memory bandwidth, core count, context length and concurrency all change results. The reliable number is the one measured on your hardware with your workload; we can help you run the same benchmark. See Evidence & evaluation for how each published claim is scoped.

Security and access

Because you deploy Helix, access control, network exposure, logging and retention are under your control and are your responsibility. Place the endpoint behind your own authentication and network controls; do not expose it directly to the internet. See security information for more, or ask us about a specific requirement.

Evaluation and licensing

Helix is licensed per deployment following an evaluation. A typical path is: try the public demo, discuss your models, hardware and boundary requirements, benchmark in your environment, then license. There is no self-service download. Start a deployment conversation.