Background
Orpheus is Canopy Labs' open source speech LLM, released in March 2025 under the Apache 2.0 license. It is built on a Llama 3.2 3B backbone trained on more than 100,000 hours of English speech, and it showed that an open model can compete with closed source rivals on prosody and empathy. It supports zero shot voice cloning, tag based emotion control, and real time streaming at roughly 200 ms time to first audio on optimized hosts. Teams can self host it or run it through managed providers such as Baseten.
Sources: github.com, canopylabs.ai
At a glance
The open weights entry on the Index: a Llama 3.2 3B backbone, zero shot cloning, emotion tags, and promised 1B, 400M, and 150M variants. Canopy Labs selected Baseten as its preferred inference provider in May 2025, with an FP8 implementation for production serving. It has no hosted first party API, so the Index shows no measured latency.
Sources: baseten.co
Position in the rankings
Standings as of Jun 13, 2026, 00:15 UTC
Frequently asked questions
- How is Orpheus tested on the Humanness Index™?
- Listeners hear Orpheus against another model in a blind head to head round, both voices reading the same customer support prompt from the same cloned source voice, and they pick whichever sounds more human. Its Humanness score derives purely from those votes.
- Is Orpheus really open source?
- Yes. The code is Apache 2.0 and the weights are published under the Llama 3.2 Community License, so teams can self host it or run it through managed providers such as Baseten. It has no commercial per character price.
How human does your model really sound?
The benchmark is open source. Suggest a model, read the methodology, or ask us to put your voice in the arena.
