VioP
2.9K posts
@AcousimHss
the more you laugh the more you cry the more you cry the more you laugh
Today, we closed our latest funding round with $122 billion in committed capital at an $852B post-money valuation. The fastest way to expand AI’s benefits is to put useful intelligence in people’s hands early and let access compound globally. This funding gives us resources to lead at scale. openai.com/index/accelera…
Today, we are emerging from stealth and launching PrismML, an AI lab with Caltech origins that is centered on building the most concentrated form of intelligence. At PrismML, we believe that the next major leaps in AI will be driven by order-of-magnitude improvements in intelligence density, not just sheer parameter count. Our first proof point is the 1-bit Bonsai 8B, a 1-bit weight model that fits into 1.15 GBs of memory and delivers over 10x the intelligence density of its full-precision counterparts. It is 14x smaller, 8x faster, and 5x more energy efficient on edge hardware while remaining competitive with other models in its parameter-class. We are open-sourcing the model under Apache 2.0 license, along with Bonsai 4B and 1.7B models. When advanced models become small, fast, and efficient enough to run locally, the design space for AI changes immediately. We believe in a future of on-device agents, real-time robotics, offline intelligence and entirely new products that were previously impossible. We are excited to share our vision with you and keep working in the future to push the frontier of intelligence to the edge.
This scatter plot shows the Pareto frontier of intelligence vs. size, defined by models like Qwen3 0.6B, 1.7B, 4B, 8B, and Ministral3 3B. The 1-bit Bonsai family shifts that frontier dramatically to the left. This changes the tradeoff itself: models no longer have to be large to be capable.
Top Text-to-Speech (TTS) Models in 2026 -- There are a ton of text to speech models out there and it's hard to know what to choose. I created some tricky text samples for 10 different models (some proprietary, some open source) to synthesize. And then I compare them for accuracy and for realism. Proprietary models are ahead here for sure, although on realism, most models today are excellent. Timestamps: 0:00 Introduction to TTS evaluation with four-row dataset and three metrics 1:08 Tricky TTS dataset: symbols, abbreviations, nouns, and prosody challenges 2:13 Prosody examples: snoring sounds, hissing, and paralinguistic elements 4:17 Roundtrip CER methodology: TTS output transcribed back with ASR 6:12 Two evaluation metrics: roundtrip CER and mean opinion score (MOS) 7:12 Results: proprietary models (Gemini, GPT-4o, ElevenLabs) achieve 4.2-4.3 MOS 10:40 Gemini demo: handles symbols and prosody but produces unexpected Irish accent 13:39 GPT-4o mini paralinguistics test: snoring example and symbol errors 15:41 ElevenLabs struggles with technical content and Irish pronunciation 17:11 Kokoro performs well but mispronounces "WV" with incorrect pauses 19:28 Orpheus model tested on unfamiliar words, Irish, and technical citations 21:30 Piper TTS quality issues: airy and choppy delivery, CPU vs GPU tradeoffs 23:12 Voxtral autoregressive model stops early with premature end token 25:15 Chatterbox produces garbled output with high CER (0.86) but realistic sound 26:15 Recommendations: Kokoro best open-source option, normalization needed for technical text 27:04 Dataset and evaluation tools available on Trelis platform
🚀 LongCat-AudioDiT is here — SOTA diffusion TTS built directly in waveform latent space. Key Highlights: ✦ Non-autoregressive diffusion-based TTS that directly operates in waveform latent space to reduce compounding errors. ✦ Simple but powerful Wav-VAE + Diffusion pipeline that achieves SOTA voice cloning performance (0.818/0.797 SIM score on Seed-ZH/Seed-Hard) and competitive intelligibility among all models (open and closed-source). ✦ Available in 1B / 3.5B, supporting high-fidelity, multilingual (ZH/EN) audio generation. ✨ Technical Innovations: ✦ By introducing the APG algorithm to replace CFG, AudioDiT improves the perceived naturalness and acoustic quality of synthesized audio. ✦ Resolves the long-standing and previously unsolved training-inference mismatch in diffusion-based TTS models. ✦ Systematic experiments reveal a non-intuitive relationship between latent space reconstruction quality and overall TTS performance: better VAE ≠ better TTS. 📄 Tech Report: github.com/meituan-longca… 🐙 GitHub: github.com/meituan-longca… 🤗 Hugging Face: huggingface.co/meituan-longca…
The artist vs the art
🚀 LongCat-AudioDiT is here — SOTA diffusion TTS built directly in waveform latent space. Key Highlights: ✦ Non-autoregressive diffusion-based TTS that directly operates in waveform latent space to reduce compounding errors. ✦ Simple but powerful Wav-VAE + Diffusion pipeline that achieves SOTA voice cloning performance (0.818/0.797 SIM score on Seed-ZH/Seed-Hard) and competitive intelligibility among all models (open and closed-source). ✦ Available in 1B / 3.5B, supporting high-fidelity, multilingual (ZH/EN) audio generation. ✨ Technical Innovations: ✦ By introducing the APG algorithm to replace CFG, AudioDiT improves the perceived naturalness and acoustic quality of synthesized audio. ✦ Resolves the long-standing and previously unsolved training-inference mismatch in diffusion-based TTS models. ✦ Systematic experiments reveal a non-intuitive relationship between latent space reconstruction quality and overall TTS performance: better VAE ≠ better TTS. 📄 Tech Report: github.com/meituan-longca… 🐙 GitHub: github.com/meituan-longca… 🤗 Hugging Face: huggingface.co/meituan-longca…
Today we’re releasing prime-rl v0.5.0. This is a major release, with 200+ commits from 22 contributors since v0.4.0. on the menu: * PD-DisAgg inference to boost agentic RL training * support for GLM-5, Qwen3.5, and Nemotron * a complete revamp of environment execution for better performance * first-class multi nodes slurm support directly from the config * quack kernel, selective AC, and more we also added several new guides to the docs, including large MoE agentic training guides. and that’s alongside many more bug fixes and improvements
