Next week (of 11th) I will be in New York. I would love to catch up with my (old) fintech colleagues and anyone deep into AI. Please DM me and I’ll arrange a meet-up one evening. Week of 18th in the Bay Area.

Why not just buy some Mac Minis and use the Qwen3-Coder-30B model on MLX? You have a more scalable cost model.

If this is OpenAI’s new OS model it’s going to be tough to run on anything but a top spec machine. It will be interesting to see (if it’s real) how it stacks up to the Qwen3 models. Either way, it will be welcome.

Asteroids, fully functional, 107 toks/sec, over 600 lines in under a minute on a laptop. Qwen3-Coder-30B-A3B-Instruct-4bit on MLX. This is the worst it's ever going to be!.

Spot on Wolfram and nice work!.

I agree with the points, especially #6.

A few months back OpenAI talked about an open model. Qwen3 has not only filled the top spot, they’ve dominated thinking and instruct with 235B and 30B LLMs along with translate and coding models too. Several releases a week, true AI innovation.

Same here Eric, this may be another nail in their coffin along with GLM 4.5, Qwen3 and Kimi K2. I’ve been a huge fan of sonnet and Opus but things move fast in AI.

Results seems to be consistant, I'm not sure I'll be using this much on my laptop but the fact that I can is what's key. This is the worse it will EVER be!

Thanks to some background research from @ivanfioravanti, the excellent MLX work of @Prince_Canuma, the incredible model from @JustinLin610 and team at @Alibaba_Qwen, I got one of the world's most powerful LLMs running on my 128GB M4 laptop. Qwen3-235B-A22B-Instruct-2507-3bit-DWQ

So far, three groundbreaking models in a week and another tomorrow. This model alone will help 95% of the world’s population talk to each other (for free). Probably one of the greatest gifts to humanity in a while!.

Another “wholly shit” day as Alibaba Qwen announce Qwen3-Coder. Just a day after they announced their Kimi-K2 quashing monster. This is another large model, the majority of us are no going to be running locally. However, you can, it’s totally open source. This new model, as the

Last week saw Kimi K2 not only take the open source crown but it crushed almost all of the OpenAI and Anthropic models. A week later the crown is taken back again by Alibaba’s Qwen3 with a new 235B model, ¼ the size of Kimi. Anyone can run these with total privacy, no one gets

I’m about to take my 18th train in Germany this year, 2nd today, 5th this week. Not one single train has been on time with delays ranging from 15 minutes to 2+ hours, averaging around 50 minutes. Deutsche Bahn is a joke, it makes British Rail look like Swiss trains.

I’m back in Munich, the organisers of @mlconference ask me last minute if I could teach a 2-day AI course. Stupidly I said yes and then spent all weekend writing exercises. Monday done, time for a beer 🍺

About to join Daniel’s rather excellent AI Happy Hour, somehow I manage to forget a drink though!! 🍺.

🤣🍺.

My colleague @AidfulAI found a prompt that seems to break the @ollama version of deepseek-r1 (8b-0528-qwen3) for Q4 and fp16. It works on MLX and base qwen3:b4. It’s…. “If a plant that doubles its size every day covers a lake in 30 days, how much time will it take for 2 plants.

For anyone wanting to update all their @ollama models in one line. ollama list | awk 'NR>1{print $1}' | xargs -n1 ollama pull.

Please join me on Tuesday evening (21h EU/ 20h UK time) for a chilled out discussion on everything AI, specifically open-source, local & private LLMs. This is run by @AidfulAI and we usually get @WolframRvnwlf on the space too.