Releases·Tool Release
Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost
Nari Labs released Qwen3-TTS and Qwen3-ASR serving endpoints backed by an open-sourced inference engine purpose-built for multimodal speech models. On Coval voice AI benchmarks, its TTS endpoint ranks #1 in accuracy and #2 in latency while being the cheapest option, beating ElevenLabs, Cartesia, and Alibaba's official endpoints. This argues open speech models can be served faster and cheaper than closed ones with the right inference stack.
119 engagement1 sourceMon, Sep 14, 2026, 04:07 PM
AI Overview
- Nari Labs open-sourced an inference engine for Qwen3-TTS running at sub-50 ms latency at 10 RPS, arguing vLLM/SGLang suit multimodal inference poorly
- Nari's Qwen3-TTS endpoint ranks #1 in accuracy (WER) and cheapest on Coval voice benchmarks vs 11Labs and Cartesia, with #2 latency
- Nari's Qwen3-ASR endpoint posts lowest latency and #2 accuracy, 0.1% behind first place, as second-cheapest on the Coval list
- Nari Labs says Alibaba's official Qwen endpoints trail its own on accuracy and latency; roadmap includes diarization, video, and world model inference
Related
Releases·Tool Release
Open-source Qwen3-TTS 1.7B hits 34 ms TTFA, 10 RPS on H100
A team open-sourced an optimized Qwen3-TTS 1.7B implementation achieving 10 requests per second and sub-50 ms p95 time-to-first-audio on a single H100, with real-time playback. The project targets local inference speed, claiming significant gains over vLLM-Omni and SGLang-Omni.
A team open-sourced an optimized Qwen3-TTS 1.7B implementation achieving 10 requests per second and sub-50 ms p95 time-to-first-audio on a single H100, with real-time playback. The project targets local inference speed, claiming significant gains over vLLM-Omni and SGLang-Omni.
1 engagement1 sourceredditFri, Aug 21, 2026, 04:01 PM
Products·launched
Latency-optimized Qwen3-ASR 1.7B streaming ASR service
A service that optimizes the Qwen3-ASR 1.7B model to achieve latency comparable to Deepgram for streaming voice calls, without changing weights. It leverages the model's inherent accuracy and multilingual capabilities, making it suitable for real-time transcription with high intelligence.
A service that optimizes the Qwen3-ASR 1.7B model to achieve latency comparable to Deepgram for streaming voice calls, without changing weights. It leverages the model's inherent accuracy and multilingual capabilities, making it suitable for real-time transcription with high intelligence.
1 engagement1 sourceredditTue, Aug 25, 2026, 03:54 AM
Releases·Model Release
Quansys AI launches Atlas-v1 speech-to-speech voice model with 0.32s responses at ~1¢ per interaction
Quansys AI introduced Atlas-v1, a speech-to-speech voice model built in-house at its Noida team. The company claims the model matches frontier reasoning quality while answering in 0.32 seconds at a cost of roughly one cent per interaction. It matters to practitioners building low-latency, low-cost voice AI who currently rely on frontier-lab offerings.
Quansys AI introduced Atlas-v1, a speech-to-speech voice model built in-house at its Noida team. The company claims the model matches frontier reasoning quality while answering in 0.32 seconds at a cost of roughly one cent per interaction. It matters to practitioners building low-latency, low-cost voice AI who currently rely on frontier-lab offerings.
0 engagement1 sourcerssTue, Oct 13, 2026, 09:16 PM
Releases·Model Release
Qwen releases open-weight TTS model but keeps top-performing variant behind API
Qwen released Qwen3-TTS with open weights under Apache 2.0, but its higher-performing Qwen-Audio-3.0-TTS-Plus is only available via Alibaba Cloud's API with no weights release. The closed variant tops Artificial Analysis's TTS leaderboard, beating Gemini and ElevenLabs, sparking community debate about Qwen's commitment to open source.
Qwen released Qwen3-TTS with open weights under Apache 2.0, but its higher-performing Qwen-Audio-3.0-TTS-Plus is only available via Alibaba Cloud's API with no weights release. The closed variant tops Artificial Analysis's TTS leaderboard, beating Gemini and ElevenLabs, sparking community debate about Qwen's commitment to open source.
46 engagement2 sourcesredditWed, Jul 22, 2026, 04:16 AM
Social·Community
Curated open-source TTS reference organized by license to highlight commercial viability
A Reddit user published an awesome-list for text-to-speech models, ordering open-source entries by license rather than quality because several popular open-weight models (F5-TTS, MaskGCT, IndexTTS-2, Higgs Audio v2) use non-commercial or usage-capped licenses that block production deployment. The list aims to help practitioners pick models they can actually ship commercially.
A Reddit user published an awesome-list for text-to-speech models, ordering open-source entries by license rather than quality because several popular open-weight models (F5-TTS, MaskGCT, IndexTTS-2, Higgs Audio v2) use non-commercial or usage-capped licenses that block production deployment. The list aims to help practitioners pick models they can actually ship commercially.
5 engagement1 sourceredditTue, Jul 21, 2026, 02:31 PM