Open models that listen: nine take audio, but only three answer in speech

Gemma 4, Apertus 1.5, Inkling, MiMo-V2.6, Muse Spark Open, Nemotron 3 Nano Omni, Qwen3-Omni, MiniCPM-o 4.5 and LFM3-Omni compared by audio input, speech output, size and license.

  • Analysis
  • 4 min read
  • Published

Nine open-weight language models in our catalog now take audio as well as text and images. Only three of them, Qwen3-Omni, MiniCPM-o 4.5 and LFM3-Omni-2B, can answer in speech. The other six write text only, and they run from an 8B model to one with about a trillion parameters. So “listens” means very different things from one model card to the next.

The nine at a glance

Model Audio-capable sizes Speech output Context License
LFM3-Omni-2B (Liquid AI) 2.17B, 4 GB in BF16 Yes 32K LFM Open License, free below US$10M revenue
Gemma 4 (Google) E2B, E4B and 12B No 128K for E2B and E4B, 256K above Apache 2.0
Qwen3-Omni (Alibaba) 30B, 3B active Yes, 10 languages 65,536 positions in the config Apache 2.0
MiniCPM-o 4.5 (OpenBMB) 9.37B Yes, English and Chinese 40,960 positions in the config Apache 2.0
Nemotron 3 Nano Omni (NVIDIA) 31B, 3B active No 256K NVIDIA Open Model Agreement
Apertus 1.5 (Swiss AI) 8B and 70B No 262K Apache 2.0
MiMo-V2.6 (Xiaomi) Flash 310B, Pro 1.02T No 1M Tagged MIT, no license file
Muse Spark Open (Meta) 816B, 42B active No 1M Muse Model License
Inkling (Thinking Machines) 975B and Inkling-Small 276B No Up to 1M Apache 2.0

Listening is not the same as transcribing

The cards limit audio in different ways. Inkling takes WAV files of up to about 20 minutes, and NVIDIA’s model takes up to an hour of audio, English only, with word-level timestamps. Gemma 4 gives audio only to its three smaller models, and its 26B and 31B models take text and images. Apertus describes its audio understanding as experimental, and its report says specialized speech models stay ahead on English and other high-resource languages.

For plain transcription, a dedicated model is still the safer choice. Our speech-to-text guide and the speech recognition list cover them.

Only three talk back

Qwen3-Omni speaks in 10 languages, but only its Instruct checkpoint has the talker, and Alibaba’s vLLM serve route returns text only. MiniCPM-o 4.5 can hear and speak at the same time in English and Chinese, and it can start a comment on its own while it watches a video stream. LFM3-Omni-2B answers in text or speech and fits in 4 GB, but its context is only 32K. For every other model, a spoken reply needs a separate text-to-speech model, as Thinking Machines says outright for Inkling. The text-to-speech guide lists options, and cloning a real voice needs the speaker’s consent.

Hardware spans two orders of magnitude

  • Phone or laptop: LFM3-Omni-2B and Gemma 4 E2B and E4B.
  • One GPU: MiniCPM-o 4.5, where OpenBMB lists 19.0 GB in BF16 and 11.0 GB in int4, Gemma 4 12B, Apertus 1.5 8B at 18.4 GB, and Nemotron 3 Nano Omni, where NVIDIA lists a 32 GB RTX 5090 for NVFP4, a 48 GB L40S for FP8 and an 80 GB H100 for BF16.
  • An 80 GB GPU or more: Qwen3-Omni, where the README gives theoretical BF16 minimums of 78.85 GB for 15 seconds of video and 144.81 GB for 120 seconds.
  • A server: Apertus 1.5 70B at 144.6 GB, MiMo-V2.6 Flash at 178 GB, Muse Spark Open at 533 GB even in NVFP4, MiMo-V2.6 Pro at 573 GB and Inkling at 594 GB in NVFP4. Inkling-Small is the practical sibling at 171 GB in NVFP4.

The licenses differ more than the sizes

Gemma 4, Apertus 1.5, Inkling, Qwen3-Omni and MiniCPM-o 4.5 are Apache 2.0. NVIDIA’s agreement allows commercial use. Liquid’s license stops being free at US$10 million of annual revenue. Meta’s Muse Model License requires the “Built with Muse Spark Open” credit and a separate agreement for hosting above US$20 million of revenue. MiMo-V2.6 is tagged MIT on Hugging Face, but its repositories held no license file on 8 November 2026, so ask Xiaomi to confirm the terms in writing. AI licenses explained covers the general rules.

How to choose

  • A voice assistant that watches a camera: MiniCPM-o 4.5. Pick LFM3-Omni-2B only if 4 GB matters more than context and the revenue cap does not apply to you.
  • Spoken answers in 10 languages, with an 80 GB GPU: Qwen3-Omni, which is more than a year old, so test it against the newer models on your own audio.
  • Long English recordings or video turned into text or JSON: Nemotron 3 Nano Omni.
  • Phones and small devices: Gemma 4 E2B or E4B.
  • The most capable open weights, with a cluster to run them: MiMo-V2.6, Muse Spark Open or Inkling. Thinking Machines says plainly that Inkling is a base for customization, not the strongest model overall.
  • A model with everything published, from data preparation to training code: Apertus 1.5, with its experimental audio.

What to keep in mind

Every benchmark table on these cards comes from the developer, and the rivals differ from card to card, so we do not rank the nine by score. This list covers the models in our catalog and not every open omni model. Test a handful of your own recordings before you commit.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
Apertus 1.5 Swiss AI Initiative 8B – 70B 256K Apache 2.0 (Permissive) 5.52 GB Laptop or phone
Inkling Thinking Machines Lab 276B – 975B 41B active 1M Apache 2.0 (Permissive) 188 GB Big-memory desktop
MiMo-V2.6 Xiaomi 9.4B – 1.02T 42B active 1M MIT (Permissive) 6.49 GB Laptop or phone
Muse Spark Open Meta 816B 42B active 1M Muse Model License (Conditional) 563 GB GPU server
Nemotron 3 Nano Omni NVIDIA 31B 3B active 256K NVIDIA Open Model (Permissive) 32 GB Gaming GPU
Qwen3-Omni Alibaba Qwen 31.7B – 35.3B 3B active 64K Apache 2.0 (Permissive) 22 GB Gaming GPU
MiniCPM-o 4.5 OpenBMB 9.37B 40K Apache 2.0 (Permissive) 11 GB Laptop or phone
LFM3 Liquid AI 1.73B – 2.17B 256K LFM Open 1.0 (Conditional) 1.19 GB Laptop or phone