inclusionAI released two native audio-language Mixture-of-Experts models on Hugging Face: Ming-3.0-audio-flash at 126B total parameters (6.5B active) and Ming-3.0-audio-tiny at 9.3B total (2.7B active). Both accept audio plus text instructions, return text or optional synthesized speech, and cover speech recognition, audio reasoning, and perception of sounds, scenes, and music. MIT licensing and documented Transformers, SGLang, and vLLM-Omni deployment paths make them usable for local voice-assistant builds.
Ant Group's lab inclusionAI released Ling-3.0-flash, a 124B-parameter hybrid-reasoning MoE that activates only 5.1B per token, under a plain MIT license. The lab claims it matches or beats its own trillion-param flagship Ring-2.6-1T. A smaller 8B variant, Ling-3.0-tiny, followed on Aug 10.
Ant Group's lab inclusionAI released Ling-3.0-flash, a 124B-parameter hybrid-reasoning MoE that activates only 5.1B per token, under a plain MIT license. The lab claims it matches or beats its own trillion-param flagship Ring-2.6-1T. A smaller 8B variant, Ling-3.0-tiny, followed on Aug 10.
Adventists AI released Sophon-Realtime v1, a family of Apache-2.0 32B full-duplex voice models on Hugging Face that listen continuously, take turns naturally, and produce expressive speech with tone. Three builds (Tag, Flash, E2E) share one brain and differ in how replies become sound. It claims top open full-duplex scores on VoiceBench, Big Bench Audio, and paralinguistic benchmarks, and is served on a single 80 GB GPU via vLLM-Omni.
Adventists AI released Sophon-Realtime v1, a family of Apache-2.0 32B full-duplex voice models on Hugging Face that listen continuously, take turns naturally, and produce expressive speech with tone. Three builds (Tag, Flash, E2E) share one brain and differ in how replies become sound. It claims top open full-duplex scores on VoiceBench, Big Bench Audio, and paralinguistic benchmarks, and is served on a single 80 GB GPU via vLLM-Omni.
MiniMax's H3 model ecosystem is growing, with a new community index cataloging integrations from local ComfyUI setups on 24GB VRAM to enterprise-scale SGLang and vLLM-Omni serving deployments. The index tracks hardware and software support across the ecosystem.
MiniMax's H3 model ecosystem is growing, with a new community index cataloging integrations from local ComfyUI setups on 24GB VRAM to enterprise-scale SGLang and vLLM-Omni serving deployments. The index tracks hardware and software support across the ecosystem.
FireRedTeam released FireRedAudio, a general-purpose audio language model built on a shared 9B-parameter LLM with decoupled continuous representations for understanding and generation. The model is available on HuggingFace with code and demos, enabling tasks like speech recognition, synthesis, and audio understanding in one framework.
FireRedTeam released FireRedAudio, a general-purpose audio language model built on a shared 9B-parameter LLM with decoupled continuous representations for understanding and generation. The model is available on HuggingFace with code and demos, enabling tasks like speech recognition, synthesis, and audio understanding in one framework.
Researchers introduce Mizar, a 159.3M-parameter audio-language model built from a CED-Small audio encoder connected to SmolLM2-135M via a frequency-merging mapper. A three-stage training recipe on ReasonAQA, AudioMCQ, and AVQA data pushes it past all prior ALMs under 200M parameters on MMAU, MMAR, and ADQA-clean. Code and checkpoints are public, and the model runs on-device with roughly one-second CPU latency.
Researchers introduce Mizar, a 159.3M-parameter audio-language model built from a CED-Small audio encoder connected to SmolLM2-135M via a frequency-merging mapper. A three-stage training recipe on ReasonAQA, AudioMCQ, and AVQA data pushes it past all prior ALMs under 200M parameters on MMAU, MMAR, and ADQA-clean. Code and checkpoints are public, and the model runs on-device with roughly one-second CPU latency.