I’m often reminded of this letter from Lead Edge capital Companies choose which data to present -- ranging from "cash profits" to "being voted the best place to work in our city."

📣 SwiftWasm 6.1 is now available Notable improvements: • No custom patches needed - fully upstreamed • Swift SDK-only distribution • Code coverage support • swift-testing support • VSCode support Check out the release blog post 👇 blog.swiftwasm.org/posts/6-1-rele…

Swift 6.2 "officially" supports WebAssembly! Let’s celebrate this milestone together 🥳 apple.com/newsroom/2025/…

1/ Today, we're proud to announce that Stripe is acquiring Privy. We couldn’t be more excited. Privy will continue as an independent product – but now we’ll move faster, ship more, and serve you even better, so you can stay focused on your users.

Edge inference demands blazing-fast structured generation. Our technology delivers exactly that and with grammars we can eliminate JSON overhead for even faster function calling. Excited to announce our collaboration with Liquid AI bringing this to edge devices!

We’re excited to announce that XGrammar has partnered with Outlines! 🎉 XGrammar is now the grammar backend powering Outlines, enabling structured LLM generation with higher speed. Check out Outlines — an amazing library for LLM structured text generation! 🚀

New on the blog: our first-ever community guest post! Baris Guler explores running AI agents fully in-browser, no APIs, no servers, powered by: ⚙️ WebLLM 🧱 WASM 🔁 WebWorkers Multi-language, local-first, and privacy-respecting. Full post: blog.mozilla.ai/3w-for-in-brow…

Interesting

🧠 [Primer] On-device Transformers • od-xformer.aman.ai - On-device transformers bring the power of LLMs and Encoder Models directly to mobile and edge hardware, overcoming constraints of latency, privacy, and connectivity while enabling real-time intelligence. - This

🚨 Apple just released FastVLM on Hugging Face - 0.5, 1.5 and 7B real-time VLMs with WebGPU support 🤯 > 85x faster and 3.4x smaller than comparable sized VLMs > 7.9x faster TTFT for larger models > designed to output fewer output tokens and reduce encoding time for high

GPT-OSS uses MXFP4 quantization (which MLX now supports). There are two FP4 formats circulating right now: MXFP4 and NVFP4 (NV for Nvidia). From looking at how GPT-OSS uses MXFP4, it is somewhat suboptimal. I'm thinking NVFP4 will be the more commonly used format in the

MXFP4 support for openai/gpt-oss landed in MLX! ⌘ + Shift + R to update your runtime. 🚀

EXO Gym: simulate large-scale distributed training experiments on a single MacBook

My next blog post is dropping this week, and it’s a much deeper dive than usual. I’ll be walking through how I fine-tuned Microsoft’s Phi-3-mini-4k-instruct (3.8B) with LoRA on my Mac using MLX. The experiment: exploring whether a 3.8B model that runs locally can be fine-tuned

real, do all the research deliberately before doing a startup

Releasing a little project I've been working on: OpenChat, an open-source macOS app to run AI models locally.

Good morning with a banger resource on vLLM

VS Code adds support for custom OAI-compatible endpoints This a big win for local AI as it allows us to use any local model provider without vendor lock-in. Big thanks to the VS Code devs and especially Isidor Nikolic for listening to the community feedback and adding this option!

🧵 We acquired OpenPipe! OpenPipe's ART framework makes it easy to beat foundation models with smaller open models on real problems, using reinforcement learning.

The The Browser Company just signed a merger agreement to be acquired. We will remain independent. Our focus is Dia. I’ve written and rewritten this post more times than I’d like to admit, but what I keep coming back to is simple: the work continues, and we’re grateful for this