SPEV: Omnimodal Jev for audio and more

Today we're releasing SPEV, the first open-weight omnimodal decision model.

Where Jev and other Jev-style models like OpenAI's Decisions and Cloudflare's Clef are multimodal, supporting text or images in, decision out, SPEV natively supports audio and video inputs. These native capabilities allow SPEV to reason directly from audio, capturing all of the nuance and detail carried in speech which would otherwise be lost during transcription, and enabling audio capabilities like agent turn detection.

#Why omni-in

Audio and video applications are some of the highest value and most time sensitive AI applications. Unlike a text-in model which has to wait on a lossy transcript, an omnimodal model can reason directly about raw audio and video.

Latency. There's no transcription step before the decision. In our benchmarks SPEV returns an audio native decision in 166 ms compared to 752 ms from OpenAI after including a transcription roundtrip.

Context. Tone, pauses, and intonation are lost during transcription. SPEV hears and understands them, allowing it to answer more difficult questions like turn detection.

#Built on Qwen3-Omni for Audio, Video, and Images

SPEV is built on Qwen3-Omni, the same base as Symphony, our speech-to-speech model. It can handle the hardest audio decisions and is available now on our API. See the Decisions docs for the full request format.

{
  "model": "spev",
  "input": "",
  "input_audio": {
    "format": "wav",
    "data": "BASE64_ENCODED_AUDIO"
  },
  "questions": [
    {
      "type": "choice",
      "name": "turn",
      "instructions": "Has the most recent speaker yielded the floor or are they continuing their turn?",
      "choices": [
        {
          "value": "keep_listening",
          "description": "The speaker is continuing their turn."
        },
        {
          "value": "respond",
          "description": "The speaker has yielded the floor; another speaker can respond."
        }
      ]
    }
  ]
}

#Benchmarks

Measured on livekit/eot-bench.

#Open Source and Open Weight

The weights are on Hugging Face. You can self-host, though it takes some setup today. The easiest path is the hosted endpoint on Sprag, which is free through October 2026.

Download SPEV on Hugging Face and tell us what you build with it.