Simon Willison
60.3K posts
Simon Willison
@simonw
Creator @datasetteproj, co-creator Django. PSF board. Hangs out with @natbat. He/Him. Mastodon: https://t.co/t0MrmnJW0K Bsky: https://t.co/OnWIyhX4CH
San Francisco, CA Katılım Kasım 2006
5.6K Takip Edilen152.1K Takipçiler
@natolambert Did we ever get a conclusive answer as to if their top researchers quit or were fired?
English
Qwen is irreplaceable. Has been going from strength to strength in recent times.
Things will always be different, I'm hopeful we can find groups of other models to fill the void.
RIP
Kevin S. Xu@kevinsxu
Alibaba said nothing about open source as part of its future AI strategy in its earnings call I thought it would at least pay some temporary lip service to open source Qwen, as we know it, is dead
English
Dan found that the 2-bit quantization broke tool calling but upgrading to 4-bit (at 4.36 tokens/second) got that working
Dan Woods@danveloper
@simonw You bet. Literally, "tool calling" became the metric that got us back to Q4. Q2 was really great conversationally and very capable, but it's like running the model at temperature 10,000 for anything predictable.
English
Wrote a bit more about this on my blog simonwillison.net/2026/Mar/18/ll…
English
Dan says he's got Qwen 3.5 397B-A17B - a 209GB on disk MoE model - running on an M3 Mac at ~5.7 tokens per second using only 5.5 GB of active memory (!) by quantizing and then streaming weights from SSD (at ~17GB/s), since MoE models only use a small subset of their weights for each token
Dan Woods@danveloper
English
@danveloper Have you observed a meaningful difference between Q4 and Q2 either when it comes to tool calling? Would love to see how you measure that
English
And now, I'm done with MoE's on this project forever. There probably is room to get to 6-8 tok/s, which even at 4 tok/s it's very usable for agentic tasks that are not time sensitive, and Q4 weights make the agent tool calls predictably reliable. Qwen 3.5 is an excellent model.
English
Some very meaningful progress on this project. A bunch of performance experiments and we've landed at 4.4 tok/s on the distribution Q4 weights. Feels pretty good since we started at 0.28tok/s. Code and experiments are up in the github repo now!
Dan Woods@danveloper
English
Thoughts on OpenAI acquiring Astral and uv/ruff/ty simonwillison.net/2026/Mar/19/op…
English
@FixTechStuff1 That doesn't matter in this case because it's effectively a read-only workload - all if that read activity shouldn't hurt the SSD at all
English
@simonw One problem with hammering your SSD like this is SSD’s have a finite number of writes.
This is fine if SSD’s are cheap and replaceable, but when it’s hard soldered to your Mac mini, then you’ll eventually have to replace the whole thing.
English
MLX is an astoundingly great piece of software which helped make Mac hardware credible as a platform for running LLMs
I'm surprised Apple didn't move heaven and earth to keep Awni, can't wait to see what what he does next
Awni Hannun@awnihannun
Today is my last day at Apple. Building MLX with our amazing team and community has been an absolute pleasure. It's still early days for AI on Apple silicon. Apple makes the best consumer hardware on the planet. There's so much potential for it to be the leading platform for AI. And I'm confident MLX will continue to have a big role in that. To the future: MLX remains in the exceptionally capable hands of our team including @angeloskath, @zcbenz, @DiganiJagrit, @NasFilippova, @trebolloc (and others not on X). Follow them or @shasha for future updates.
English
@simonw Empirical with Opus doing the sanity checking. I’m not sure 2-bit quantization even mattered that much in the end… it was an earlier test, so I’ll probably revert that and see how it does with regular 4-bit. The k=4 was a binary search by Claude, checking the quality each time.
English
@NoeFlandre SVG is a little more useful, I actually have models produce a SVG for real web features sometimes
English
Notes and pelicans for today's GPT-5.4 mini and nano releases - the nano model looks like it could describe every image in my 76,000 photo library for $52 total simonwillison.net/2026/Mar/17/mi…
English
@ClementDelangue Qwen 3.5 was shockingly good, even at tiny sizes like the 4B model (which somehow benchmarks similar to GPT-4o across many of the classic benchmarks)... and then much of the core Qwen team quit or were fired (still not clear to me which) straight after releasing it
English
@cyrusradfar @GergelyOrosz "In the end we're all communicating"
Not if our AI assistant made the decision to reply to something and then wrote and posted a reply
English
In the end we're all communicating.
I'm not clear on why the effort put in matters. It's whether the message comes through. To discount the message because it was AI supported, feels unfair.
Tech leaders don't write their social posts, but we don't say "they didn't write that" -- we just take it , respond or react as we like. The method of creation is irrelevant.
I get that we all don't want slop, but that existed well before AI online -- especially in online "discourse."
English
It’s not X — it’s Y
I cannot unsee how so much of the writing on this site (and online, in general) is increasingly AI-generated. It’s still pretty easy to recognize. Probably not for long tho
Just alarming that ppl outsource even typing 3 sentences for a reply on this site…
English
@ryanjanssen Given how good their Claude Code for web hosted version is I would be shocked not to see a hosted Claude Cowork from them soon
English
@simonw these are all bandaids on their main problem (that CC is natively local)
I’m interested how they’ll address the underlying need for cloud
English
Couldn't resist getting OpenAI Codex to render me a pelican for every combination of model and reasoning effort - I do think gpt-5.4 xhigh came out the best, the pelican has a fish in its beak!
English
... and a follow-up chapter about Subagents, now a feature of Codex and Claude Code and Gemini CLI and Mistral Vibe and OpenCode and VS Code and Cursor simonwillison.net/guides/agentic…
English
New chapter for Agentic Engineering Patterns: I tried to distill key details of how coding agents work under the hood that are most useful to understand in order to use them effectively simonwillison.net/guides/agentic…
English