🇧🇪 lived in 🇨🇭🇦🇹🇸🇰 now 🇩🇪 Med AI Democratisation 25+ yrs Digital Health leadership, father ex-IBM ex-SAP @isaree_ai https://xitter.cf/t.co/TsBCgl9Ldm eu/acc
Berlin & Antwerp
Joined November 2007
- Tweets15.1K
- Following2.4K
- Followers7.4K
- Likes10.1K
Bart de Witte retweeted
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months
Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
Bart de Witte retweeted
At this point the whole EU Inc negotiations are so lost in the details that everyone forgot the main goal: Creating a standard.
Basically, allow founders to raise quickly globally and use stock options to hire competitively.
Let’s say it quite clearly:
1) If EU Inc isn’t providing a standard it’s useless. We already got 30+ legal entities in Europe. Experienced founders use Delaware Inc already because of this.
2) If EU Inc doesn’t fix stock options Delaware Inc is still the better option.
I understand the there are complex details and people are worried about loopholes. But you can’t lose the actual goal of it because of it.
An EU Inc that doesn't provide a standard is pointless. Nobody will care about intricacies of compromises reached in political committees if the market can't use the outcome.
If you are in contact with your political leaders (prime ministers, presidents, et al) please reach out to them. Or to founders/investors in your network who might have that access.
🔥🇪🇺✌️
xitter.cf/euinc_petition/status/…
Bart de Witte retweeted
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
I agree with @elonmusk but we saw this coming 2 years. 3 scenarios (6/4/3-month doubling) with the shaded uncertainty band and the base-case milestones marked: laptop-class agentic capability in 4 GB by end-2027, clinical orchestrator plus subagents co-resident by end-2028, 2026 cloud-frontier routine capability in a 4 GB phone model by 2030. ago. 1/3
All first principles have been going into one direction. Log-scale MMLU points per GB for 23 models, with the best-of-year frontier (Gopher, Chinchilla, Phi-2, Qwen2.5-1.5B) and the fitted 4.6-month doubling line, plus the deployment-precision series (Q4 and BitNet native) on top showing the compounding effect of compression. The footnote flags why the panel closes after 2024.2/3
GB needed for a fixed capability level per year, with device memory reference lines (4 GB app budget, 7 GB iPhone usable, 16 GB laptop). GPT-3-level capability falls 875x, from 350 GB to 0.4 GB, crossing the phone budget line in 2023-2024. This is the chart that makes the @isaree_ai thesis visible.
The quiet story of 2026: GPT-3-level intelligence went from 350 GB of datacenter hardware to 0.4 GB on a phone in under five years. The next platform shift is not up there. It is in your pocket. We have been building for that future.
Bart de Witte retweeted
Muse Spark 1.3 will be open weight! That’d make it the best open weight model if it releases before next GLM
Meta is finally bringing back open weight dominance to the US
Bart de Witte retweeted
Very proud to announce the first episode of
@EUTechLoop radio with Bart de Witte
@OpenMedFuture. We talked about:
▪️open source + open weights
▪️the second leg of the AI race
▪️the AI Omnibus + the Digital Omnibus
▪️and, of course, the #EUinc.
eutechloop.com/bart-de-witte…
I touched on this in my opening @openEHR keynote last October: we are moving from mainframe AI to personal AI ten times faster than we transitioned from IBM mainframes to the Apple Macintosh.
Qualcomm CEO reveals AI companies want phones running 100-billion-parameter models continuously by 2028--and confirms some are building phones
Cristiano Amon: “I have to be careful not to disclose names, but some AI companies—some of the companies at the forefront of AI—are telling Qualcomm, ‘By 2028, I need to make sure that I have the ability to run at least a 100-billion-parameter model in a phone, and I need to have this thing running kind of all the time.’”
Alex Heath: “So you’re saying there are AI companies that are talking about—”
Cristiano Amon: “Well, AI companies are starting to do phones.”
Bart de Witte retweeted
Apple is currently incapable of moving quickly, so I hope Ternus is making a lot of changes. Playing with Omarchy has taught me that an AI native OS can be VERY compelling, and I am sure an AI native phone would be as well.
Qualcomm CEO reveals AI companies want phones running 100-billion-parameter models continuously by 2028--and confirms some are building phones
Cristiano Amon: “I have to be careful not to disclose names, but some AI companies—some of the companies at the forefront of AI—are telling Qualcomm, ‘By 2028, I need to make sure that I have the ability to run at least a 100-billion-parameter model in a phone, and I need to have this thing running kind of all the time.’”
Alex Heath: “So you’re saying there are AI companies that are talking about—”
Cristiano Amon: “Well, AI companies are starting to do phones.”
What do we want call healthcare stakeholders who rely on LLMs trained on open data, but will not contribute data that does not legally belong to them?
They have seen how quickly shared data accelerates AI innovation, and in healthcare, that translates into lives saved. MIMIC-IV should not still be the only valuable EHR dataset out there.
This is insanely good news!
A coordinated $1.8 billion commitment from Biohub, the Department of Energy, NIH, Google DeepMind, Isomorphic Labs, and Meta to generate accessible multimodal data and measurement technologies for predictive AI models of the cell is exactly the kind of foundation the field has needed.
The real compounding value, as with the Human Genome Project, will come from making the data usable by the broader research community so models of cellular behavior can be built, tested, and improved by many groups at once.
That kind of infrastructure shortens the path from molecular measurements to better disease understanding and, ultimately, to treatments.
The emphasis on open access is the part that matters most.
This is insanely good news!
A coordinated $1.8 billion commitment from Biohub, the Department of Energy, NIH, Google DeepMind, Isomorphic Labs, and Meta to generate accessible multimodal data and measurement technologies for predictive AI models of the cell is exactly the kind of foundation the field has needed.
The real compounding value, as with the Human Genome Project, will come from making the data usable by the broader research community so models of cellular behavior can be built, tested, and improved by many groups at once.
That kind of infrastructure shortens the path from molecular measurements to better disease understanding and, ultimately, to treatments.
The emphasis on open access is the part that matters most.
🧵 Biohub, @ENERGY, @NIH, and new funding partners today announced a major expansion of an international effort to generate and make accessible the data enabling predictive AI models of biology. Together, the organizations are investing $1.8 billion in funding, data, computation and new measurement technology - largest coordinated commitment to map the biology of the cell to power development of digital models of biology. bit.ly/3UhjoxD
Bart de Witte retweeted
Mistral Large 4 just tied #1 on 669 clinical decisions 🔥
617 right on day 0, level with Jev, Clef and pplx-decider. the only model at the top with zero severe misses: no missed emergency, no allergy read as a medicine taken.
Le Chonk can triage 😅: 238/240, the best so far!
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
Le Chaton Fat is here!
🇪🇺 still delivers!
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
Great to see this. Perhaps some hope that soon we can use the famous quote from Gary Lineker. @grok can help to understand the reference.
thanks for sharing! a bit more context:
* we’re not a random german lab. Aleph Alpha was founded in 2019 and launched its first model in 2022.
* not SOTA open-weight, but very strong on the quality-cost frontier. yes, we are ahead of Qwen3.6, Mistral Small, Nemotron Nano and Super. Qwen3.8-27B has better overall quality, but has ~8× more active parameters.
* yes, we trained from scratch on ~24T tokens, with more than 20% German in pre-training. that’s more German tokens than the major open German datasets combined!
* we own the data pipeline, training and evaluation stack and built the infrastructure and processes to iterate quickly and continuously improve (stay tuned for what comes next).
you can find many more details in the paper. grab a coffee (or two) and get ready for 189 pages on how we did it :)
aleph-alpha.com/downloads/te…