Lucas Newman

@lllucas

i like building things. prev: apple and startups, now working on something new.

San Francisco, CA

Joined April 2007

Replies Reposts Media

Lucas Newman
@lllucas
Aug 26
day 0 support for GLM-5.3-Flash and Qwen3.8-Flash-Next now available! Nativ is open source, free, hackable, and built in SwiftUI so it feels right at home on your mac.
Nativ
@Nativ_AI
Aug 26
Congrats to @Zai_org on GLM-5.3-Flash! Day 0 support in Nativ 🎉 320B params, 18B active · 1M native context · fully on-device. On an M3 Ultra (512GB) with Nativ v0.3.4, standard 4-bit MLX: ⚡ Up to 505 tok/s prefill · 32 tok/s decode 🚀 Batch 4 lifts decode to 70.9 tok/s 💾
Lucas Newman
@lllucas
Jun 12
maybe we should rename it mlx-omni? 🤔
Ivan Fioravanti ᯅ
@ivanfioravanti
Jun 12
MLX-VLM: wait, what? Support TTS and STT added by @lllucas 👀 mlx-vlm is becoming mlx-anything 🔥
3
Lucas Newman
@lllucas
Jun 8
CoreAI looks like a nice upgrade from CoreML, with more practical coverage for modern LM workloads like dynamic input shapes and better mutable buffer handling for kv caches. The tooling is still early but excited to see where it goes. The beta release has Pytorch model
1
Lucas Newman
@lllucas
Feb 23
something I've wanted to exist for a long time: an easy way for iOS and macOS apps to have access to the latest speech technology on-device! we've built a developer-friendly API with state of the art models like Qwen3 TTS, plus the latest in speech-to-text, diarization, and
Prince Canuma
@Prince_Canuma
Feb 23
Day 1 of 3 days of MLX: Introducing MLX-Audio-Swift SDK 🚀 A modular Swift SDK for voice agents and tasks on Apple Silicon built by @lllucas and yours truly. iOS, macOS, and visionOS developers can now build native apps with real-time, on-device audio intelligence: 🗣️
00:00
3
Lucas Newman
@lllucas
Feb 16, 2025
Ever wanted to build your own text-to-speech model? I made Nanospeech, a simple, modern, and hackable text to speech system with PyTorch and MLX implementations in ~1500 LOC. It has a tiny pretrained model (82M params) to try out, generates speech at 3-5x realtime on modern
3