Magnus Lundstedt (@MagnusLdt) header
Magnus Lundstedt (@MagnusLdt) profile picture

Founder & CEO, Precisit

Stockholm, Sweden
Joined October 2012
Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)Photo by Magnus Lundstedt (@MagnusLdt)
Jensen stuck his neck out for defending open models, knowingly took a huge risk and paid a high price and “saved the world”. It all makes sense now. Hopefully his efforts were enough.
1
47
Every day there is a new interesting layer to this story
Interesting: the Washington Post says that Jensen Huang's downfall was Marco Rubio's doing (washingtonpost.com/technolog…), and it looks like the "spy" framing was directly his and not the CIA's. As they write, Rubio took a CIA report and rewrote it in a memo "in his own more assertive terms." Which fits neatly with CNN reporting that the CIA is "not yet convinced" that Alice Huang was a spy (edition.cnn.com/2026/11/12/p…): if the CIA is not convinced, yet a memo rewritten from the CIA’s own report presented her as a spy, then there’s only one person who could have turned "not convinced" into "spy": the man who did the "more assertive" rewriting, aka Marco Rubio 🤷 The WaPo article also mentions the motive: China hawks like Rubio had spent years trying to "destroy the influence Jensen had over the president", and WaPo calls his ouster "a huge victory" for them as it allows them to "seize on the moment to redirect the national agenda toward tougher policies on China." And, as a matter of fact, they’re already pushing three new chip-restriction bills into the defense bill now that Jensen is no longer around to lobby against them. If true, it means the real "influence operation" targeting Jensen was run out of Washington, not Beijing 🤷
61
Looking forward to this! 💯🚀 I expect Qwen 4 flash will be a fantastic local model on 128 GB mac
🚀 Qwen4 is coming! Discover the latest in AI at Alibaba Cloud Summit Hong Kong 2026. Explore how Qwen4’s deeper knowledge, richer world understanding, and stronger coding and agentic capabilities are designed for real-world tasks — from complex coding and knowledge work to front-end development and professional software workflows. Join us for a full day of practical AI, enterprise innovation, and industry collaboration. Register now: moxie-mail.com/project/Aliba… #AlibabaCloud #Qwen #AlibabaCloudSummit #HongKong
2
1
95
Of course.. 🤷‍♂️🤦‍♂️
News: The US intel community is still working to understand the relationship between a Chinese businesswoman who had an affair with Nvidia CEO Jensen Huang and the Chinese government, with multiple sources telling CNN that the intelligence collected on her to date does not prove that she was acting as a spy for Beijing. cnn.com/2026/11/12/politics/…
2
26
Sad news from Nvidia today, this is a massive setback and huge risk for open weights models, increasing the risk of regulatory capture of AI by large corporations. All of us in the open source AI community have to carry ideals forward and work hard to keep the lights on, continue sharing ideas and building together. Be prepared for a possibility that Huggingface may not be the best platform in the future.
1
50
This was a fun project 😁🤓
We trained an AI to play Connect Four in a single forward pass, with no search. Quantized to ternary weights (−1, 0, +1), the model is 1.6 MB, and our own 37 KB WebGPU engine picks a move in about a millisecond, right in your browser.
25
Magnus Lundstedt retweeted
The becoming begins.
1,730
1,469
1,379
18,756
2,913,136
Looking forward to qwen 4 🥳
Qwen4, coming soon… CTO quote was “next week” Xudong Guo - Research Scientist, Qwen, Token Foundry Live from Amsterdam
53
Cool
So one of our ambassadors @divagr1925 has been working on something cool He found a way to use Qwen's Gated DeltaNet state as memory, and built a 32-bit RISC-V CPU directly into the model's weights. So instead of relying on predicting the answer to a calculation, the model can execute instructions internally. It can also run C and MicroPython programs. He calls it Qwenputer, love the name btw. The results so far are pretty impressive. On one arithmetic benchmark, accuracy jumped from 48.6% to 100%. Pretty cool to see developers experimenting with Qwen's architecture. Great work, my friend!
8
Practicing dancing? Loop any part of a song directly in the browser. Open, click share -> add desktop shortcut to keep persistent local storage. loops.dance/
26
Samma modell, ny motor.​ Vi skrev om runtime till WebGPU. Tid för AI drag gick från 13 ms till 1 ms och koden från 14 MB till 37 kB. Samma 8 MB vikter. Körs i webbläsaren. precisit.github.io/onepass-w…
1
26
Great step in the right direction! Would be nice with more permissive licensing for use in products. Not just research.
Apple dropped new model LensVLM-9B on Hugging Face and it Beats RAG-style baselines up to 10.1x. It’s a 9B Qwen3.5 finetune that turns long documents into compressed page images (5-15x), scans them visually, then expands only the relevant pages back into full text. - Render the document as tiny page images - Scan the compressed pages - Expand only the pages that actually answer the question - Near full-text accuracy at 4.3x compression. Result: token-efficient long-context QA with near full-text accuracy at 4.3x effective compression. Paper says it beats RAG and other baselines up to 10.1x. - huggingface.co/apple/LensVLM…
21
With Mac Studio being mostly used for local AI by owners, Apple should invest in their own AI lab for releasing SOTA open weight models
2
1
48
Enkel sida för att följa diff i rösträkningen i valet 2026 rörande uppsamlingsdistrikt samt slutlig rösträkning med data från valmyndigheten (uppdateras var 30 min) magnuslundstedt.github.io/el…
1
194
We've gone full circle with vertical dock from NextSTEP to iOS on iPhone Duo Compare with the original layout: youtube.com/watch?v=rf5o5liZ…
37
A few days ago Astra was the hottest model, but right now I am most excited for DeepSeek v4.1 flash! AI speed running singularity? 🤣 Available for testing with DeepSeek API directly, using model ”deepseek-v4.1-flash-expires-on-0910” 🥳
79
DeepSeek V4.1 Flash release tomorrow September 10. Pricing seem to be reduced as: Input, cache hit: 0.007/0.014 -> 0.003/0.006 (/1M) Input, cache miss: 0.22/0.44 -> 0.15/0.30 (/1M) Output: 0.66/1.32 -> 0.6/1.2 (/1M) Depending on traffic, could see savings of 15-40% 🎉
78