Ant Group's Ming-3.0 Beats Grok and Gemini on Speech Reasoning With 99.8%
Ant Group's open-source Ming-3.0-audio-flash scores 99.8% on Artificial Analysis's speech reasoning benchmark, ahead of Grok Voice 3.0, Gemini 3.8 Live and GPT-Realtime-2.1. The weights are MIT-licensed.
Takeaways
- Ming-3.0-audio-flash claims #1 on Artificial Analysis Big Bench Audio speech reasoning at 99.8%.
- It edges out Grok Voice 3.0, Gemini 3.8 Live and GPT-Realtime-2.1 High, all closed models.
- MoE model with 126B total parameters, 6.5B active per step; MIT-licensed weights on Hugging Face and ModelScope.
- Leads vendor ASR charts, especially Chinese dialects; trails rivals on meetings, long audio and tool calling.
- Full speech serving needs eight 80GB GPUs; SGLang-Omni runs on two; no hosted API yet.
- Lighter 9.3B Ming-3.0-audio-tiny runs on a single GPU for local use.
Ant Group’s inclusionAI lab says its open-weight Ming-3.0-audio-flash model ranks first on Artificial Analysis’s Speech Reasoning leaderboard. Its 99.8% Big Bench Audio score exceeds Grok Voice 3.0, Gemini 3.8 Live Extended Thinking (High), and GPT-Realtime-2.1 High. The downloadable, MIT-licensed model therefore leads paid real-time voice APIs on one narrow test: answering logic puzzles delivered as synthetic speech. It has yet to receive a complete voice-agent score, and its 126B weights make self-hosting a multi-GPU job.
126B weights, 6.5B active
Ming model card describes a native audio-language model that accepts audio and text instructions. It supports speech recognition, spoken question answering, audio understanding, audio reasoning, and optional synthesized speech. Its mixture-of-experts language backbone contains 126B total parameters while activating 6.5B for each inference step. The active count reduces computation per token. Memory planning must still account for the full set of experts.
The five-model Ming-3.0 release covers speech synthesis, audio understanding, and speech recognition under the MIT license.
| Model | Primary role | Total parameters | Active parameters |
|---|---|---|---|
| tts-0.5B | Speech synthesis | 0.5B | 0.5B |
| tts-tiny | Speech synthesis | 7.9B | 1.3B |
| audio-flash | Audio understanding | 126B | 6.5B |
| audio-tiny | Audio understanding | 9.3B | 2.7B |
| asr-tiny | Speech recognition | 8.6B | 2B |
Inside the audio path
Ming routes audio through an encoder that converts sound into embeddings, followed by a projector that maps those embeddings into the language model’s input space. The MoE backbone then processes the acoustic representation alongside text instructions. An optional Talker component converts the resulting output into speech, allowing one model stack to cover recognition, reasoning, and synthesis without requiring an intermediate transcript.
The configuration lists 512 experts, eight active experts per token, 42 layers, and a maximum position setting of 131,072 tokens. The architecture combines Multi-head Latent Attention with Kimi Delta Attention to control the memory and compute cost of long sequences. It also uses multi-token prediction, which trains the model to predict several upcoming tokens at once. The 131,072-token setting represents the model’s post-encoding context budget; usable audio duration depends on encoder compression and the serving configuration.