This is very impressive, on the leaderboard with a 15B, 128k context reasoning model. I shall be testing this today, nice work guys!

This will be right up Prince Canuma’s street for MLX!

Tuesday at Devoxx Belgium, James Ward, Rod Johnson and I had a workshop on Agentic AI using Embabel. I have my last talk later this afternoon on local private LLMs.

Full house today at Rod Johnson’s Embabel talk at Devoxx Belgium.

John Davies giving an important talk on local LLMs. Currently describing how Incept5 has built a banking workflow on Embabel using only local models. Java John T Davies 🇪🇺 Incept5

Travelling back from a great week at Devoxx. I love travelling by train, this is the relatively new Frecciarossa service from Paris to Marseille. Working with a full meal service and an office at 320km/h (over 200mph), cheaper than a flight and very civilised!

I fed this image (below) from Prince Canuma's ongoing work on MLX-VLM to support Qwen3-VL and batch. Qwen's Qwen3-VL-30B-a3B-Instruct-4bit (MLX) with the following prompt. For me on an M4 it was... Prompt: 895 tokens, 711.575 tokens-per-sec Generation: 134 tokens,

This will be incredibly useful in parallel agentic processes. You rock Prince!

They should make a Breaking Bad episode on this guy, he sure can cook!!!

So it’s just a souped-up Raspberry Pi. Get a 48 or 64GB Mac Mini for half the price and it will run the same models way faster, mine runs the same qwen3-32b-8bit at over 30tps and that’s not even with MLX.

Apple M5, the pin to burst the bubble? linkedin.com/posts/jdavies_…

Prince, king of MLX joins the big-boy club!

OK, I'm now browsing on ChatGPT-Atlas, no way is it getting access to my machine but so far, it's pretty cool.

Just downloaded and the MLX version (lmstudio-community/Qwen3-VL-2B-Instruct-MLX-bf16), great image decoding and over 100 tokens/second. Next the 32B versions.

Back in Germany, all 3 trains were late, even the one we took because we missed the other. However, they do serve beer on the train!

310 Tokens / second with an unquantised (bf16) model using MLX, this is CRAZY fast! I've just tried the new MLX-VML beta from Prince Canuma (using an M4 Max). DeepSeek-OCR (3B) on MLX is a game-changer.

Excellent news for Ollama users, one of the most powerful models out there is (finally) supported by the Ollama team. Nice work guys!

This looks very interesting, a perfect size too... moonshotai/Kimi-Linear-48B-A3B-Instruct On MMLU-Pro (4k context length), Kimi Linear achieves 51.0 performance with similar speed as full attention. Offering significant speedups at long sequence lengths (1M tokens).

In stunning Krakow this evening. Hoping to catch up with Prince Canuma tomorrow.

Wow!!! We spoke over lunch for 6 hours, so many ideas, Prince Canuma is going to take over the AI world. Remember this day folks, it started in Krakow! Top secret for now, co-investors, apply now!