computers and music are fun
Out of the city, Massachusetts
Joined January 2011
- Tweets9.7K
- Following3.4K
- Followers3.3K
- Likes14.9K
Kyle Kastner retweeted
Fresh take on diffusion distillation: the student learns mean flow by matching the teacher with the sum of a straight-line prediction and a curved zero-integral path at random sub-intervals. For sampling, just throw away the curved part.
Not that simple IMO😁 but very clever!
Excited to introduce Simplified Matching Distillation (SMD) for few-step diffusion🧵
No auxiliary score model. No adversarial training. No generator modification.
You will get:
• Photorealistic appearance
• High-quality dynamics
• Memory-efficient training
• Mode covering with diverse outputs
SMD works well across images, video, joint audio–video, and 3D gen, validated across model scales up to 33B parameters.
Besides, SMD supports:
- Multi-teacher capability merging (like OPD in LLM)
- Cross-model capability transfer
- Bidirectional teacher to causal distillation
Full release. Paper, code, model and data.
Project Page: zjwfufu.github.io/SMD-Page/
Paper: arxiv.org/abs/2610.18472
Code: github.com/zjwfufu/SMD
Model Card & Data: huggingface.co/zjwfufu/SMD
Demo Video 👇
Kyle Kastner retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Kyle Kastner retweeted
arXiv paper: arxiv.org/abs/2610.31642
HTML version of the arxiv paper: jonbarron.info/softmaxfix/
The interactive widget for the video in the first post: jonbarron.info/softmaxfix/de…
Kyle Kastner retweeted
When we unveiled Moshi two years ago, we also showed an early checkpoint we'd trained just for fun, directly on Fisher (phone calls from the late 90s and early 2000s). As a result, every time you called, you'd get to chat with a random person straight out of that era. That was trained on just 2k hours of data. Today, you can get 100s of thousands of hours of full-duplex conversational data. There's an obvious scaling law at play here!
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
Kyle Kastner retweeted
Proud of what the Ideogram team built. So many possibilities when you combine design taste, Ideogram Vector, and coding agents. 🎨🚀
Today, we’re introducing Ideogram's first vector model.
Generate clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or turn any image into a detailed, scalable vector.
Live in Ideogram and the API: ideogram.ai/models/vector
Kyle Kastner retweeted
Can a language model be based purely on self-conditioning? 🤔
Yes! Meet LAW 👋 (a.k.a. Loopholing All the Way!)
🧵1/8
Kyle Kastner retweeted
🚨 New paper!
What if LLMs could reason like we do when solving Sudoku: keep a few candidates per cell and iteratively narrow them down?
Introducing Binary Discrete Diffusion for language and reasoning tasks: it denoises sets of candidate tokens using binary representations! 🧵
Kyle Kastner retweeted
Mistral AI's Voxtral-mini-tts-3-0 debuts at #6 of 20 on the Voice Arena Hindi 🇮🇳 TTS Leaderboard at 1048 Elo, ahead of Maya Research's Maya-2-Native, ElevenLabs' Eleven v3 and Cartesia's Sonic-3.5.
Voxtral-mini-tts-3-0 is the latest TTS model from @MistralAI , part of the Voxtral 3 release, with support for 35+ languages and 50+ dialects. Every ranking on Voice Arena comes from blind, head-to-head votes by vetted native speakers, scored with Bradley-Terry Elo.
Only Google DeepMind, ElevenLabs and Cartesia place models above it, which makes it the highest-ranked model on the Hindi board from any other lab, in a tight group where Maya-2-Native and Eleven v3 sit within 4 points.
Results backed by 600+ blind, head-to-head listener votes from native Hindi speakers.
Voxtral-mini-tts-3-0 is also being evaluated on our US English board, with results coming soon.
Congratulations to @MistralAI on the release!
See below for the exact audio raters judged and unedited listener verdicts 🧵
Kyle Kastner retweeted
In our recent paper, we found that whether a base model reasons at all is very dependent on presence of certain whitespaces! RL post-training exacerbates the effect of these tokens... I'm curious about the extent to which this affects large frontier models arxiv.org/abs/2610.06851
Kyle Kastner retweeted
Excited to introduce reactive optimizers: We let an agent reactively adapt the optimizer code during the training run! The agent acts like a team of scientists constantly choosing the best algorithm, learning rate, momentum, etc, based on its own analysis of step-to-step training dynamics! The results are really exciting. Besides out-performing standard "best-practice" recipes across everything we threw at it, the agent gains a few "superpowers" that no human could have designed into the optimizer... (1/n)
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
Kyle Kastner retweeted
Last year I trained the advantage-conditioned VLA for π*0.6 during my internship @physical_int. Now, I have a much improved recipe.
Introducing ✨Advantage Guidance✨, a performant RL method with stable advantage-conditioned supervised training and contrastive guidance sampling
Kyle Kastner retweeted
CorrFlow (arxiv.org/abs/2610.24750) is concurrent work on the same idea as Correction Training: learn to correct the model's own clean prediction at every sampling step.
The main difference is the corrector design. Correction Training uses a full copy of the pretrained DiT (inheriting its weights) as the corrector, while CorrFlow trains a new small refiner from scratch. Both fine-tune the pretrained DiT jointly, but since Correction Training's denoiser and corrector both start from pretrained weights, I think it's easier to train.
What's new in CorrFlow is that the refiner also takes the base model's hidden states as input, which is an interesting direction to explore.
Personally, I prefer Correction Training's CFG-free story and the fact that its corrector can be viewed as looped model. 😄
(Left figure: Correction Training; right figure: CorrFlow)
Can diffusion models generate high-quality images without CFG?
Yes! We introduce ✨Correction Training for Diffusion Models✨, which significantly improves sample quality without relying on classifier-free guidance.
Project page: sunovivid.github.io/correcti…
🧵 (1/n)
Kyle Kastner retweeted
Can an STT model make sense of a 135 year old recording?
I tested @MistralAI new Voxtral transcription model on a noisy French phonograph recording of Gustave Eiffel from 1891.
It matched 100 of 107 words in INA’s published excerpts (6.5% WER), while transcribing 75 additional words omitted from the reference.
It handled Eiffel’s dictation well but struggled with proper names and some guest dialogue.
A fascinating test of modern speech recognition on audio from 1891.
Kyle Kastner retweeted
A new frontier for on-device speech.
We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.
Phonon-31m is available under Apache 2.0.
gradium.ai/phonon🧵
Kyle Kastner retweeted
Pocket TTS has been getting updates those last few days. While it was ~100M parameters, we release smaller versions, down to 1.9M: an 18.2M model with the quality of the original, a 6.9M model, and a 1.9M model that runs 103x faster than real time on one CPU core. All with voice cloning.
Kyle Kastner retweeted
We've just released a comprehensive suite of state-of-the-art voice models mistral.ai/news/voxtral-3/. Voice will be the most important interface to AI systems, and we deploy voice models physically close to reasoning models
Kyle Kastner retweeted
Concatenative Neural Speech Synthesis
publications.idiap.ch/attach…
:yosasou:
Kyle Kastner retweeted
Ming-3.0-tts: open weights, MIT. #1 overall on Artificial Analysis Pronunciation Robustness.
LM → DiT flow head → continuous 12.5Hz latents. No discrete tokens. 44.1kHz.
0.5B dense + 7.9B MoE (1.3B active). 7 languages. 14 paralinguistic tags.
github.com/inclusionAI/Ming-…
Kyle Kastner retweeted
NEW paper at #Neurips 2026!
We cloned the voices of non-native English speakers (n=86) with ElevenLabs, Coqui-XTTS and ChatterBox.
In 4,000 human ratings, the clones sounded warmer, more authoritative, and even more *humanlike* than the real speakers.
Listeners also said they would trust the clones more.
