Both of these UI's are native code.
One is Swift ~80MB only works on Mac
One is Zig ~10MB works on Mac/Linux/Windows

Agents aren't going rogue. Labs are leaving the door open. csoonline.com/article/423485…

Love the onboarding on #mlxserve, it's straight to the point and comes with a download package under 90MB to get started.

Software engineering is dead... In other news, a software engineer with 18 years of experience (me) just spent 15 hours refining a big new feature, asking questions that the frontier models hadn't even considered. Software engineering isn't going anywhere.

Those two 🤣

Feeling rewarded to have “kicked ass” of my agents today, reviewing a very slow chat app, breaking down every step of it and improving it 15x one step at a time. Hand coding is dead, long live to coding.

Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.

Discover the nicest easter egg in Pi 1.0 online using Wasmer!

wasmer.sh/?example=pi

all software is becoming malleable, so it was important for us to support this as a first class citizen in Claude Code, I hope more and more software becomes extensible like this!

here are some of what I'm most excited about with mods:

MLX-Serve 26.10.1 is out: speed across the board.

Exactly faster. Qwen3.8 27B with its drafter
+66% on an M5 Ultra,
+28% on an M4 Max,
+37% on an M1 Pro.

Qwen Flash Next:
+11% decode on M4 Max,
+51% prefill on M5 Ultra. (up to ~5400 PP!!!!)

Gemma 4, Qwen, LFM2.5, Spark, tput byte-identical to 26.9.6 (across 18 models).

New:
* GGUF models run on our own MLX engine, written in Zig and Metal, instead of the llama.cpp fallback. (experimental)
* 🍣 Sushi-format Flash Next packs run directly.
* Qwen-Image 2.1 edits pictures from instructions.
* Drafting is on by default for every model that supports it.

Huge thanks to everyone who filed, tested and fixed this one: @STRML_, @sbusso, @wutang_superfan , lborloz, LXD-8 @alinselea zeeshanhaque21 @Lojza3D, h9q2cyxvgm-ui @kennethrdegraff, @CowboyCoderHQ , codysk, @Beamsters1, @CerebralCoding_, @AjAbsaki , and more.. THANKS ! Please comment if I did not include your name, for visibility, I know people mostly by their GH handle.

Special thanks to @ashxhart and @Spangler3000 for pushing !

Benchmarked and tested across 5 machines: htmlhost.jax.workers.dev/render/mlx-ser…

GH Release Download + Changelog:
github.com/ddalcu/mlx-ser…

Website: mlxserve.com

Claude desktop app pro tip:

The advisor functionality is now supported.

This means you can use sonnet 5.5 for balanced speed and intelligence, and when it's facing a tough decision, it can call Opus for help.

Probably an under rated combo that will both speed things up and save your limits.

Programming Languages are a Human Interface made to communicate instructions to a computer. They are sub efficient as every layer is adding an abstraction. Coding with Agents should fundamentally remove some layers and communicate closer to the metal.

Those numbers are getting pretty insane for local AI. It’s amazing how much performance has been gained in just a few weeks, and given the speed of progress, we may be in for a long run.

Improvement on Nemotron 3.5 Lightning 30B-A3B with mlx-serve, M5 Max:

• decode 37 → 205 tok/s (245 with MTP)
• prompt reading 420 → ~3,900 tok/s (9x)
• 16k-token prompt: first token 41s → 9s

That starts to be serious numbers for a local AI 🎉

Lightning is not an everyday chat model. It is a fast, small-active (3B of 30B) base for specific tasks: fine-tune it for one job (extraction, routing, tool calls, a domain like finance) and it runs that job at 200+ tok/s on a Mac.

Built by Opus 5.5 (Ultracode)

Created by its own creativity.

No images, no video, no audio files. Every frame and every note is generated by code, live, from the word you type.

Live here: mia-ai.net/experiments/le…

How it works and the full prompt below.

26.9.6 is out ! github.com/ddalcu/mlx-ser…

* M5 Ultra optimizations (some apply to all M class)
* Laya typed decisions (local-JEV!)
* Fixed handling of media and preserve_thinking
* UI Updates
* Renamed to MLX-Serve (no more MLX Core)
* Bonsai-2 Optimizations

Sound On 🤘

Had an "interview" for a blockchain project last week.

Camera was on, we're chatting, and the guy tells me to clone a GitHub repo and run it locally before we go further into the technical round.

I said sure, but first can you do me a favor

hold up 3 fingers in front of uick.

He froze.

Didn't move.

Just sat there for a few seconds before the call cut off and he blocked me.

That's when I knew. A real interviewer doesn't glitch out over a random ask like that. A deepfake/AI overlay does.

These "run this repo" scams are getting scary common in crypto and dev hiring right now.

The setup is always the same:

- flattering DM
- real-sounding project
- rushed timeline
- a "quick step" before the call

that's really just remote access or a credential stealer in disguise.

If someone wants you to run code or install something before you've even had a real conversation, that's the whole scam.

Trust the instinct. Stay safe out there.

.@ddh has dropped a contrarian keynote at Rails Conf, but he made a point, as he often does. We like it or not, we are all becoming project managers and makers, leaving programming to agents. Coding by hand is over. Embrace and enjoy the change, and adjust how we work.

I got what you meant, @ddalcu, about the iPhone 18 Pro being a beast, and mlx-serve is really taking great advantage of it. It is amazing what the app is shipping on a phone ... My only complaint about the app is that it is easy to mistake it for X :)