Homi

Homi is an open decision model that processes audio clips and questions to return option probabilities.

Homi is an open decision model that takes in audio natively along with typed questions, returning a probability for every option without generating text. Based on Qwen3-Omni-30B-A3B-Instruct with frozen weights, a decision head, and LoRA parameters, it handles paralinguistic and phonological questions that transcripts miss. It provides yes/no, choice, and score question types for tasks like audio monitoring and evaluation. The codebase includes inference scripts, evaluation runners, and training data pipelines. It is released under the Apache-2.0 license for both code and weights.

Homi at a glance

Best for: Researcher
Runs on: self-hosted, open source
Pricing: Free
Released: October 2026
Category: AI research tool
Website: github.com/egma-ai/homi

Key features

Answers typed questions about audio clips
Returns a probability for every option
Supports yes/no, choice, and score question types
Uses frozen Qwen3-Omni-30B-A3B-Instruct base weights
Includes inference code and a command-line interface
Provides evaluation benchmarks and runners
Includes training data builders and recipes

Use it for

Monitoring and evaluating large-scale audio at low cost
Answering paralinguistic and phonological audio questions
Assessing speech pitch, speed, volume, and emotion

Alternatives to Homi

Oumi — Automates model training and evaluation to cut costs and improve accuracy.
NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
Decido — Runs locally on CPU while beating published benchmark scores.
Vegaml — Runs locally on your hardware with calibrated probabilities and conformal answer sets.
LMMs-Eval — Provides open research and tooling for multimodal model evaluation and benchmarks.
OpenCompass — It offers an objective, comprehensive, and credible open-source AI evaluation system.
Ncm — Generates text significantly faster than diffusion baselines while maintaining high quality.
Vev — Runs locally on NVIDIA hardware with a fast, Jev-compatible API.
SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
Hemory — Connects real-world conversations to your AI agents via MCP.
Anime Text Dataset — Provides structured anime data with field-level provenance and question-answering pairs.
Blumify — Tracks daily interviews and speeches of prominent AI leaders.

All Homi alternatives