SCROLL NEWS / DISCOVERY
Search headlines
Find the stories shaping the conversation.
Results for " mistral " 15 found 🔀 AI-powered Shuffle
‘Close to zero’: 4 AI chiefs say no companies have truly changed in response to AI
At a Fortune event, executives from Accenture, AXA, Mastercard and Mistral AI said the technology isn't the main obstacle to AI payoffs.
Companies pledge free software and computing for AI in science
The European Commission has announced that a group of companies and foundations will offer free services, such as software or access to supercomputers, to European scientists using AI in their research.Unveiled on October 15 at a conference in Dublin, the initiative is the latest step by the Commission to try to help researchers use the technology. Scientists have used forms of AI, such as machine learning, for decades. But advances in deep learning in the past decade, and the wider interest in large language models since the release of ChatGPT in 2022, have spurred governments to invest in new AI tools that could help the scientific process.In July, the US announced it is spending $5 billion on the Genesis Mission, which it describes as a “national effort to harness AI for science,” funding 278 projects across areas such as autonomous labs or new AI tools in astronomy. The Commission has also identified AI in science as a priority, but the €107 million it pledged from Horizon Europe in 2025 does not come close to US spending levels. That said, the vast majority of public research spending in the EU comes from national funds. Today’s announcement at the European AI in Science Summit doesn’t commit any new public money. Instead, it launches the AI in Science Enablers Initiative, a set of voluntary contributions from companies and foundations. The idea is that the companies also benefit by getting access to scientific ecosystems, and feedback on how their tools work. Google, for example, has put up €3 million of cloud computing credits to European researchers. The Novo Nordisk Foundation intends to make its supercomputer, Gefion, available to European researchers, according to the Commission statement. The chipmaker Nvidia said it would provide access to tools, compute, training and collaborative projects.The publisher Elsevier also offered complementary licenses to some of its AI tools, plus AI training for PhD, postdoctoral and master’s students. The French AI firm Mistral will provide the European research community with access to its open-weight models, Vibe seats and application programming interface credits, according to the Commission’s statement. Vibe is one of the firm’s chatbot models. The idea is that more firms and foundations will come on board in the future as the initiative grows.
Mistral raises €3B as sovereign AI becomes big business
The French AI lab has raised €3 billion at a €21 billion valuation in a Series D round led by Samsung, Scaleup Europe and PSG Equity.
French A.I. Start-Up Mistral Raises $3.5 Billion as Part of Strategy Shift
Mistral is trying to keep pace with American and Chinese rivals while offering customers a European alternative for artificial intelligence.
How Attribution Science Uses Models to Uncover Climate Change’s Effects on Weather
Running climate model simulations with—and without—observed human influences can quantify how emissions changed the probability or intensity of a weather event.
How should European scientists respond to the US AI kill switch?
Earlier in June, Europe’s worst fears of technological dependence were confirmed when the White House ordered the AI firm Anthropic to stop non-US citizens using its leading models. This kill-switch moment crystalised anxieties in Brussels that the continent, lagging the US on the creation of leading-edge AI models, is now a technological of vassal of Washington. The shock has specific implications for European scientists and universities. Some governments are now worried that with AI being integrated into labs, Donald Trump now has his finger hovering over an off switch for European science. “Having restrictions on this tool, it can have an impact on our own research,” said France’s research minister, Philippe Baptiste, at the VivaTech conference in Paris on June 18.Baptiste, himself a former academic, said that colleagues in mathematics had recently warned him they were now cut off from leading-edge US models, risking their work. “This model is not available anymore now because of the US restriction, and for the first time this had an impact on mathematics,” a field where large language models (LLMs) are making a significant contribution, said Baptiste. So, should European scientists avoid relying on US models in the future? Or would this be a self-harming overreaction? And how is AI being used in science, anyway? What happened?On June 12, it emerged that Washington, citing cybersecurity concerns, had ordered Anthropic to stop non-US citizens using two models: Claude Mythos 5, and Claude Fable 5, a version of Mythos but with guardrails to prevent abuse. Anthropic said Washington’s concerns were overblown. But forced to follow the order, yet with no way to check who was using the models, the company suddenly shut down access for everyone, US citizens included. There’s still no agreement on how to restore broad access. But last week, Bloomberg reported that select US companies, and the US government, still had access to Mythos despite the ban, but the EU’s cybersecurity agency Enisa had been shut out. Class division of researchersIn the longer term, access outside the US should be restored, said Marcel Salathé, co-director of the AI centre at the Federal Polytechnic School of Lausanne (EPFL), because a permanent restriction would be financially ruinous to US AI firms. “Even Trump will eventually budge to economic pressure,” he said.However, US AI firms might end up operating rather like banks, doing due diligence on users of their leading-edge AI models, he went on.European scientists are unlikely to fall foul of such checks, he said. But having a Chinese postgraduate in their lab, or doing a type of research the US disapproved of, could potentially result in access being denied. “It will create a sort of class division of researchers,” he said. More than frontier modelsHowever, successfully deploying AI in science is about far more than just having access to leading LLMs such as Mythos.Instead, many science-specific tools, most famously Alphafold, which predicts protein folding structures, require “far less” computing power to train than frontier models, well within the capabilities of European supercomputers, according to a recent analysis by the Arq Foundation, a Brussel-based AI think tank. Instead, tools such as Alphafold are possible because certain fields, like protein folding, weather prediction or astronomy, have large, clean, standardised datasets that can be used to train predictive models. Rather than compute, “data is the central constraint for AI for science,” the report says. For example, a less powerful physics model trained on CERN’s data “will be more valuable than the best model without certain data,” said Bengüsu Özcan, author of the report. European advantage?Europe could therefore steal a march by using its scientific data and infrastructure to create models US digital giants cannot. “They might actually match or go even beyond what these big [US] companies are doing,” said Özcan.Indeed, some European start-ups hope to create automated testing labs to create the kind of standardised data that could speed up the discovery of new materials. Materials scientists say there’s currently too little data to train models. AI scientistsHowever, even if the bottleneck for AI in science is often data, rather than compute, being at the mercy of the US for LLM access could still become a problem for European scientists. A new class of “AI scientists” based on leading LLMs is also emerging, Özcan’s report says.Unlike domain-specific tools like Alphafold, these try to automate day-to-day scientific work, by providing an overarching tool that coordinates specific models to synthesise existing literature, generate hypotheses or run analysis of data. The problem for Europe, however, is that these tools are sometimes based on US leading models. Google Deepmind’s AI Co-Scientist, for example, is built on Google’s Gemini LLM. Systematic failureThese scientific automation tools are still in their early stages, and still exhibit “systematic failures on scientific tasks,” Özcan’s report says. One of the problems is that the messy, day-to-day process of scientific discovery is largely not documented, meaning LLMs don’t have a good source of data to learn how scientists actually work. Meanwhile, some tasks, such as literature synthesis, don’t typically require access to leading edge models.However, the performance of AI scientific assistants could improve as companies such as Google pour resources into developing them. Rely on US“Unfortunately, for scientific assistants, there is not much we can do right now except rely on US frontier AI,” said Antoine Bosselut, a natural language processing expert based at EPFL. Chinese-built open models, which can run on local computers, lessening the risk of a shutdown, are touted as one alternative to US dependence. There’s also Mistral, once seen as Europe’s main AI hope, but the Paris-based company has seen its recent models fall behind in performance. But the Chinese alternatives are generally not as capable as US frontier AI models, said Bosselut, unless paired with fiddly, custom tools that allows them to act as independent agents. “Most scientists do not have the time,” he said. Don’t jump ship yetAlthough the kill switch moment is seen a warning shot for European scientists, it doesn’t mean labs should abandon US AI tools immediately. “I do not think the right response is simply to abandon US-based models and systems under the banner of sovereignty or independence,” said Mattias Björnmalm, secretary general of the Cesaer university association. “That would be an overreaction.”Related articlesEuropean sovereignty in AI requires ‘ugly trade-offs,’ say experts Robotic lab start-up makes case for €500M EU-funded materials testing facilityAI in science: is it useful? Instead, European researchers need to ask themselves whether they’re becoming dependent on technology that could be turned off in Washington, said Özcan. “They might not see this today, but they should really start asking this question,” she said. European researchers need a plan B, said Salathé. “I would recommend to every researcher, very proactively explore alternatives.”Own toolsIn addition, Björnmalm and others want European research institutions to develop their own scientific AI tools.Instead of signing AI deals with US digital giants, European universities should be using this money to create their own models, which may end up being superior for research, said Sabine Leonelli, an expert on the use of AI in science at the Technical University of Munich. “Universities in Europe seem to be happy to sign commercial contracts with US companies on the assumption [their AI tools] are better,” she said. But US frontier LLMs are typically opaque, and have been developed without independent scientific input, she said. “This is not good enough for science.”Last year, Swiss researchers released Apertus, an LLM designed to be completely transparent, from its development methods to its training data. The idea is that this makes scientific tools built on top of the model far more reproducible than with closed, commercial alternatives.“Europe needs more of this: serious capacity-building that gives researchers, universities and companies credible options,” said Björnmalm.
We can create the future of science right now
Paul Litvak is the founder and Executive Director of the Robyn Dawes Institute and a Visiting Scholar at UC Berkeley. He has a PhD in Behavioral Decision Research from Carnegie Mellon where he studied emotions and the sunk cost bias. Over 15 years in industry, he solved a wide range of challenging technical problems. At Meta he created models optimizing the ad review process. At Google he ran experiments to measure social influence. At Airbnb, he was a product manager leading machine learning teams optimizing search ranking and price suggestions. He also co-founded and led product at Rhythmic Health, a venture-backed biosensing startup, creating an accurate low cost system using a color changing strip and a smart phone to measure salivary lactate. The bottleneck At this point, it is uncontroversial to say that science needs to stop using the PDF article as the unit of knowledge and currency. The unbearable slowness of scientific publishing, the profit motive and margins: I’m not saying anything new. The PDF also sucks because it’s hard to extract structured information, which makes it hard to do evidence synthesis. As a result, we do much less evidence synthesis than is needed. And evidence synthesis ultimately undergirds most policy and medical decision-making. I can see second-by-second real-time odds for any sporting or newsworthy event, but a school board can’t see the best evidence on whether their 8th graders should be taught algebra. As a society, we don’t treat this as an important problem. Again, not controversial. Not only is the problem well understood, but the solution has already been laid out. What we need is AI-assisted living evidence synthesis - (1) an open knowledge graph of atomic claims, (2) claims linked to evidence, (3) assessment and synthesis of each piece of evidence, and (4) continuous updating with new data. What few realize (yet) is that the technical capacity to build this vision for a significant portion of science already exists. Not only that, scientists and startup teams are already building many of these components. I know this because I’ve been surveying the space and talking to many of the builders. There are some missing pieces: for example, evaluations of how well some of the components work. But at this point, most of what’s missing is a fully end-to-end working integration of all of these parts. In the rest of this essay, I’m going to lay out all the parts of a working living evidence layer for science and who is working on them, and propose concrete next steps for building this system. What’s now possible The diagram above outlines the components of a living evidence synthesis platform, including some of the teams working on each component1. Scientific PDFs are processed into claims with associated evidence. The evidence is subjected to a forensic audit, methodological evaluation and robustness and reproducibility checks. Finally it’s given a weight in a continuously updating synthesis. What follows is a description and status of each component and a few of the teams working on them. Document understanding The first thing you need to be able to do is turn an article into structured data. Mostly that means parsing PDFs. There are often multicolumn layouts that confuse non-specialized PDF to text processing libraries. For scientific papers, there is the added complexity of parsing formulas and tables and figures. This is a really hot area - there are startups offering APIs, and it seems like a new open source package gets posted to Github every few weeks. What follows isn’t exhaustive. A package called GROBID was the state of the art for a while, they didn’t update their package for nearly two years until very recently. In the meantime reducto.ai released an AI powered PDF extraction API, PaddleOCR became popular, IBM released a model called Docling, and both Mistral and Gemini created models and libraries. I also know of at least one other well-funded psychology research group working on a paper parser. By contrast, there are few open evals in this space, with no extensive evals for complex table comprehension in particular. Nonetheless, I’m confident this will be a solved problem soon, given the combination of LLM advances and developer interest. Hypothesis level extraction There has been increasing interest in comprehending the extracted text of papers and linking information to evidence for each hypothesis. A lot of work has already been done. Trialstreamer (Marshall et al. 2020) and RobotReviewer LIVE (Marshall et al. 2023) demonstrated automated extraction of trial population, intervention, and outcome at scale on clinical RCTs. PaperQA2 (Skarlinski et al. 2024) and Ai2 ScholarQA (2024) extended this to retrieval-augmented question answering with citation grounding. Elicit, Consensus, and SciSpace operationalized claim-level extraction for end users. OpenEval (Booeshaghi et al. 2026) is the most recent and most ambitious: 1.96 million atomic claims extracted from 16,087 eLife manuscripts using Claude Sonnet 4.5, grouped into ~299,000 results, with LLM evaluations showing 81% agreement with human peer review on a 2,487-paper subset. None of these solutions link claims to test-statistics, as you would need to evaluate randomized controlled trials. This is why I built the evidence.guide API - to extract hypotheses and associated test statistics from behavioral science papers. The best public eval of this kind of extraction I’m aware of comes from the recent SCORE project - they had humans code thousands of psychology papers to extract their claims by hand. It would be extremely helpful to the world if all scientific PDFs were available as structured open data. I’ve been working to make this happen, both directly at Berkeley and through coordination with large entities I can’t yet speak of; as hard as it is to do, I think it’s possible2. Forensic audit A lot of work has been done on forensic audit, but some gaps remain. Of course, for biology papers that rely on images for evidence, there are a variety of tools (notably Proofig and ImageTwin) to spot anomalies. These are still well short of what sleuths like Elizabeth Bik can do on her own, but these tools are constant companions among fraud analysts. There’s someone working on auditing Excel files for anomalies, and a number of teams are automating numerical checks like GRIM and SPRITE, including Scrutiny project, the INSPECT-SR team as well as statcheck. The regcheck team is building a way to use AI to compare preregistrations to analyses in papers, to ensure there aren’t significant deviations. Nonetheless, there are many other kinds of anomalies to screen for, both public and less publicly known. And there are no formal evals for anomaly detection that I’m aware of. Still, there’s a lot to draw from in this space and I’m pretty certain we will be able to scan papers for most kinds of obvious anomalies in the near future. Methodological review This area has been white hot, though I fear for many of the startups in this space, because this capability may become commoditized. There are at least six different AI peer review companies, including Refine.ink, Reviewer3, ReviewerZero.ai, Q.E.D. Science, Paper Wizard, and Isitcredible. Coarse (a pun on refine) was also recently created as an open source alternative. These systems provide qualitative feedback on the content of papers, spotting methodological weaknesses and mathematical errors. They seem to work pretty well, and many academics report bitterly that they exceed the average quality of typical peer reviewers. But there are few evals here either. What evals exist so far involve using LLM-as-judge (circularity problems abound) or comparing against human reviews of questionable quality. What you’d ideally want is an eval that measures capturing known errors in papers3. Reproducibility and robustness Another active area has been using AI agents to automate computational reproducibility4 and robustness5 checks in papers that report numerical results. For more recent papers where data and code are available, AI agents can see whether they can re-run the analyses and produce the numbers reported in the published paper. In addition to a handful of individual academics who have been experimenting using Claude Code for this, the Institute for Replication is a leading group working on building an end to end system. The evals related to this problem are the most mature, with CORE-Bench (Siegel, Kapoor, Narayanan 2024) and PaperBench (OpenAI 2025) available to benchmark agents on this task. There is also work on getting AI agents to test alternative ways of analyzing the data to ensure the results are robust to small analytic design choices. Synthesis This is the most underdeveloped area where significant investment is required. Although some automated evidence synthesis systems exist — for example, otto-sr is building an AI agent to write systematic reviews — none of these incorporate the full range of paper level signals to weight evidence appropriately. Nor is there anything like an eval or a gold standard for a good systematic review. Arguably Cochrane reviews are the closest we have to gold standard human systematic reviews, though I’ve heard academics in the know complain about their uneven quality. A key question for a synthesis platform is how to weight anomalies and methodological issues in assessing the quality of a piece of evidence. This is an unsolved problem and one I’m very keen to work on. Continuous updating There are many pieces of basic infrastructure available for monitoring for new research and initiating updates. OpenAlex is the current open citation graph. Retraction Watch integrated into Crossref in October 2023. Scite tracks how citations support, contrast, or mention prior claims. The Living Evidence Network demonstrated continuous-update workflows in clinical guidelines. Engineering this is a relatively straightforward task. When you look over this technical architecture and all the progress being made, it’s hard not to be optimistic that a living guide to scientific evidence will be built. The stakeholders are ready The social infrastructure for this is starting to coalesce — it’s not just a pie in the sky academic exercise to imagine this coming into existence. Institutions like the Center for Open Science, the Institute for Replication, the INSPECT-SR, the Living Evidence Network and more are all working on scaling work to improve research quality. Funders are also aligned. The Sloan Foundation has funded living evidence work through COS. Coefficient Giving supports the Institute for Replication and COS. The Astera Institute and the Institute for Progress have shown interest in this space. NIH has established an Office for Replication and Reproducibility. Although there are (very unfortunately) serious headwinds in science funding generally, there is an active group of funders interested in metascience. A brief word about what I’ve been doing at RDI. First, as a Visiting Scholar at Berkeley I’ve been actively figuring out how a non-profit and a public university can conduct and make public the results of large-scale academic article data mining. With some of the money I raised from donors, I commissioned a legal analysis of recent case law and publisher text data mining (TDM) agreements in order to understand whether a massive open data mining of academic articles is possible (with caveats, it is). I’ve also been working to bring together stakeholders in this space, and identify gaps. I’ve also been doing some software development in this space, with more to come. A pilot proposal The assumption undergirding all of this is that an AI, given all this information, would make the right judgment about a scientific claim with lots of conflicting evidence, weighing all the factors appropriately. That’s the hypothesis we need to test. Randomized control trial research is the best place to focus on first. RCTs are used to make many of the important decisions in society - from medical trials to public policy changes. And they use a relatively uniform set of inferential statistics with lots of known and available diagnostics. Behavioral science experiments, within the broader realm of RCTs, should be first used as a testbed whose results can be generalized. Because behavioral science is at the vanguard of open science practices, replications abound (there are thousands of them) to serve as ground truth training data. Key Hypothesis Therefore the pilot would test, in behavioral RCTs that have been replicated, whether the quality of evidence for a claim can be used to accurately predict whether that claim will replicate. Secondary Hypotheses Compared to claims that replicated, non-replicated claims demonstrate a greater share of forensic anomalies in their source literatures. Hypothesis level claim and statistic extraction is accurate enough to scale living evidence without onerous human review costs. Replication prediction is more accurate than prediction markets6 or journal prestige. If all the different quality signals we gather do accurately predict which studies will replicate, then we can use that model to score evidence to power the living evidence layer. Why this is informative regardless of outcome If the pilot succeeds, the architecture extends to medical RCTs (where Living Evidence already operates and integration is mostly about claim representation), then to slices of basic biology with stable replication structure. If it fails, the field learns which quality signals are load-bearing and which ones metascience has oversold. Either result is a contribution to knowing what the literature supports. Implications for funders Because this burgeoning ecosystem of builders already exists, a well-informed philanthropic or government funder could play a crucial catalyzing role in bringing this future about. They could play at least three roles: creating open structured datasets, publishing open benchmarks and incentivizing the solving of key technical challenges. First, open archives of papers that the government maintains, like PubMed, could be turned into structured data amenable to large scale metascience and claim aggregation. I know the US government already has an interest in doing this, though some key open questions remain unanswered. How do you determine the best models and systems for accurately extracting information from papers? How can you establish a robust way to allow researchers to flag errors and correct them? And finally, how do you create a legal regime in working with publishers to maximize the scope of available papers? For the latter, a university or a private philanthropy may be better positioned to make structured data publicly available under journal subscription terms or fair use. Philanthropists or government funders could also coordinate to create or commission benchmarks that evaluate whether important problems have been solved. For example, an open benchmark for claim extraction from a range of different scientific article types would be extremely helpful. Ensuring the underlying data are accurate is vital for creating these evaluations. I’ve discussed opportunistically using various human-created datasets for this purpose, but a consistent problem is that errors in human data make it difficult for them to serve as a gold standard. Finally, with structured data and benchmarks available, the government or private philanthropy could use them to incentivize groups to develop machine learning systems that meet these benchmarks. Prizes are one potentially valuable tool for this. For example, you could establish a prize for a system that accurately updates a living meta-analysis for a small set of claims. Prizes are particularly useful as signals of problem importance, and can help create vibrant ecosystems of public and private research—see, for example the role the government played in kickstarting the current work on nuclear fusion. Conclusion The drawbacks of the current scientific publishing system are known. Scientists agree, metascientists agree, philanthropists agree: the published PDF plus citation graph isn’t the right substrate for maintaining a representation of the evidence base in science. The pieces needed to build the alternative either already exist or are rapidly taking shape. The community is forming around exactly this problem, with concrete partnerships and shared infrastructure. A pilot should start on behavioral science RCTs because that’s the slice of empirical science most amenable to legibility, where replication ground truth is richest, and where the failure modes are best documented. What’s been missing is the galvanizing mission to assemble these pieces into something that works. That’s what I’m proposing to build. *** 1 I have a broader field map that I’ll release publicly soon. This is me, building in public! 2 If this is something that you are excited about, please reach out and talk to me. 3 More on this very soon too! 4 This tests whether, given the code and the data, you can get the same statistics as reported in the published paper. 5 This tests whether the results are the same as a paper’s given alternative analytical decisions in conducting the analysis (like outlier omission). Closely related is the idea of a “multiverse” where you come up with many different ways of answering the same underlying research question with the same data, and test whether the results hold in all those alternative methods. There’s been work on the latter as well. 6 Some of the replications, e.g. those from the SCORE project, had paired the experiments with forecasts from prediction markets. So we get to look at this for free.
Mistral courts U.S. business with cheaper, not the smartest, AI
Hot off its first corporate showcase in Paris, the crown jewel of France’s AI industry is making a play for American banks, businesses and tech bros.
I went to Mistral's summit and heard a clear message about AI
I attended Mistral's first AI summit in Paris. Here's what executives and attendees said about Europe's AI moment.
Mistral CEO warns Europe has 2 years to avoid AI dependence on US
Mistral's CEO Arthur Mensch said AI dominance will depend on who controls chips, energy, and computing infrastructure.
Elon Musk explores Mistral partnership to take on AI rivals
Musk is eyeing a collaboration with French startup Mistral to help xAI catch Anthropic and OpenAI.
Gemma 4 and what makes an open model succeed
Having written a lot of model release blog posts, there’s something much harder about reviewing open models when they drop relative to closed models, especially in 2026. In recent years, there were so few open models, so when Llama 3 was released most people were still doing research on Llama 2 and super happy to get an update. When Qwen 3 was released, the Llama 4 fiasco had just gone down, and a whole research community was emerging to study RL on Qwen 2.5 — it was a no brainer to upgrade. Today, when an open model releases, it’s competing with Qwen 3.5, Kimi K2.5, GLM 5, MiniMax M2.5, GPT-OSS, Arcee Large, Nemotron 3, Olmo 3, and others. The space is populated, but still feels full of hidden opportunity. The potential of open models feels like a dark matter, a potential we know is huge, but few clear recipes and examples for how to unlock it are out there. Agentic AI, OpenClaw, and everything brewing in that space is going to spur mass experimentation in open models to complement the likes of Claude and Codex, not replace them. Especially with open models, the benchmarks at release are an extremely incomplete story. In some ways this is exciting, as new open models have a much higher variance and ability to surprise, but it also points at some structural reasons that make building businesses and great AI experiences around open models harder than the closed alternatives. When a new Claude Opus or GPT drops, spending a few hours with them in my agentic workflows is genuinely a good vibe test. For open models, putting them through this test is a category error. Something else to be said about open models in the era of agents is that they get out of the debate of integration, harnesses, and tools and let us see close to the ground on what exactly is the ability of just a model. Of course, we can’t test some things like search abilities without some tool, but being able to measure exactly the pace of progress of the model alone is a welcome simplification to a systematically opaque AI space. Share The list of factors I’d use to assess a new open-weight model I’m considering investing in includes: Model performance (and size) — how this model performs on benchmarks I care about and how it compares to other models of a similar size. Country of origin — some businesses care deeply about provenance, and if a model was built in China or not. Model license — if a model needs legal approval for use, uptake will be slower at mid-sized and large companies. Tooling at release — many models release with half-broken, or at least substantially slower, implementations in popular software like vLLM, Transformers, SGLANG, etc due to pushing the envelope of architectures or tools. Model fine-tunability — how easy or hard it is to modify the given model to your use-case when you actually try and use it. The core problem is that some of these are immediately available at release, e.g. general performance, license, origin, etc. but others such as tooling take day(s) to week(s) to stabilize, and others are open research questions — with no group systematically monitoring fine-tunability. In the early era of open models, the days of Llama 2 or 3 and Qwen pre v3.5, the architectures were fairly simple and the models tended to work out of the box. Some of this was due to the extremely hard work of the Llama, Qwen, Mistral, etc. developer teams. Some is due to the new models being genuinely harder to work with. When it comes to something like Qwen 3.5 or Nemotron 3, with hybrid models (either gated delta net or mamba layers), the tooling is very rough at release. Things you would expect to “just work” often don’t. I’ve been following this area closely since we released Olmo Hybrid with a similar architecture, and Qwen 3.5 is just starting to work well in the various open-source tools that need to all play nice together for RL research. That’s 1.5 months after the release date! This is just to start really investing more into understanding the behavior of the models. Of course, others started working on these models sooner by investing more engineering resources or relying on partially closed software. The fully open and distributed ecosystem takes a long time to get going on some new models. All of this is lead-in for the most important question for open models — how easy is it to adapt to specific use-cases? This is a different problem for different model sizes. Large MoE open-weight models may be used by entities like Cursor who need complex capabilities in their domain, e.g. Composer 2 trained on Kimi K2.5. Other applications can be built on much smaller models, such as Chroma’s Context-1 model for agentic search, built on GPT-OSS 20B. The question of “which models are fine-tunable” is largely background knowledge known by engineers across the industry. There should be a thriving research area here to support the open ecosystem model. The first step is to understand characteristics of different base and post-trained models to understand what they look like. The second step is to tune pretraining recipes for open models so they’re more flexible. Interconnects AI is a reader-supported publication. Consider becoming a subscriber. Subscribe For The ATOM Project and other Interconnects endeavors, we’ve put in substantial effort to measuring adoption trends in the open ecosystem. Everything takes a long time to unfold after a model is first publicly available — and adaptability is why. What we know for sure now, when Qwen has been going from strength to strength with its releases, is that technical staff across the industry has gotten comfortable working with Qwen models. Countless research methods and datasets were made to work with Qwen. It’ll take patience for any other model family to get to this point — a patience I’m not sure many open model builders have. This takes us to Gemma 4, Google’s latest open models. Gemma 3 was released more than a year ago, in March of 2025, and is a bit underrated. Gemma 4 comes in 4 sizes for now, with a bigger, MoE model of over 100B total parameters rumored but not released yet. The models we have today come in sizes of ~5B dense, 8B dense, 26B total 4B active MoE, and 31B dense. I’m most excited that they’re finally adopting a standard Apache 2.0 open source license. This’ll massively boost adoption. The standard of better licenses for strong open-weight LLMs was set by mostly Chinese open model labs in the last 1-2 years, and now U.S. companies are following suit. I will personally be so happy if the horrible Llama licenses and Gemma terms of service were an ~18-month transient dynamic of the industry being nervous about releasing strong open models. The Gemma 4 scores look very solid, the small models have incredible benchmark scores (especially in general domains like LMArena) and the 31B model rivals the recent Qwen 3.5 27B, which is the leading member of that class. The ~30B size range is an important one, as it’s accessible both to researchers and to enterprises looking to deploy the model in real use-cases. Where the 7B model scale is the default for tinkering and research, a 30B model is the default for seeing if an open model can unlock substantial value in your specific workflow — a good mix of intelligence, low price, tractability for downstream training, etc. This takes us back to the above adoption criteria I mentioned for open models and the bigger question — do I think Gemma 4 will be an overwhelming success? Previous Gemma models have been plagued by tooling issues and poorer performance when being finetuned. Gemma 4’s success is going to be entirely determined by ease of use, to a point where a 5-10% swing on benchmarks wouldn’t matter at all. It’s strong enough, small enough, with the right license, and from the U.S., so many companies are going to slot it in. I’m cautiously optimistic that Gemma 4 is going to work better here. Winds are shifting for open models built in America. We saw GPT-OSS go through a bumpy launch to become an overwhelming success. There’s a collective energy around the likes of Reflection, Arcee, Nemotron, Gemma, Olmo, and peers that show substantial demand for building new stacks around open models. There’s capital to be spent on AI stacks across the economy by those who want more ownership of everything, including the model. After launching The ATOM Project 240 days ago, the conversation is shifting into the next stage. Summer of 2025 was a crisis moment where the U.S. AI scene realized it can’t wait and figure out open models after building AGI. The two markets will capture different areas and proceed in parallel. Now that more companies in the U.S. are releasing strong models, we need to improve the ecosystem so that these models are easy to use, understand, and build value around. It’s the hard work to build another inflection point in these adoption plots I’ve been updating consistently, but that’s the work to be done. Join me in it. More data coming soon! Here’s a sneak peek:
Record-Breaking One-Off Bugatti Brouillard Could Become the Most Expensive Bugatti Ever
A one-off aimed at a historic record Bugatti’s latest one-off, the Brouillard, steps onto the hypercar stage with the quiet confidence of a record-breaker. The brand’s tradition of bespoke excellence meets a bold new aesthetic, positioning this unique commission to potentially become the marque’s most expensive creation. Built on the proven underpinnings of the Chiron/Mistral architecture, yet transformed by distinctive craft, the Brouillard is both homage and innovation. A name born from equestrian legend More than a century ago, a striking white horse captivated Ettore Bugatti in Molsheim. That equine myth now informs the car’s name and its sensory identity. The result is an automotive portrait of athletic grace—surfaces that read like muscle, curves that suggest motion, and details that quietly speak of discipline and heritage. Embroidered horse motifs on door panels and seat backs echo the story, while a miniature glass sculpture of the steed crowns the shifter. Sculptural design with purpose The Brouillard’s silhouette favors sinuous surfaces over hard creases, allowing light to glide over its form like a tendon under skin. Aerodynamics inform elegance: an integrated ducktail replaces a deployable wing, cleansing the airflow and boosting stability at speed. Its stance is athletic, its volumes taut, and its visual language unmistakably Bugatti. “Sculptural, organic forms—like a tendon beneath the skin—hide immense power behind noble simplicity,” said design chief Frank Heyl, capturing the car’s philosophy in a single stroke. Craft elevated to an art The cabin embraces craftsmanship at an uncommon level. Custom-woven textiles from Paris introduce a discreet tartan pattern, while green-tinted carbon fiber nods to the exterior’s hue. A glazed roof creates a cathedral atmosphere, bathing the cockpit in natural light. A central spine appears to flow from exterior to interior, emphasizing continuity and structure. Highlights include: Custom textiles with subtle tartan motifs Green-tinted carbon fiber matched to the exterior finish A “cathedral” glass roof that elevates cabin ambience A sculptural centerline linking outside to inside Bespoke embroidery and glass equine detail W16 power at its absolute apex Under the sculpted skin, the final evolution of Bugatti’s 8.0‑liter, quad‑turbo W16 delivers a thunderous 1,600 hp at 7,050 rpm and 1,600 Nm of torque. This represents the pinnacle of nearly two decades of Bugatti engineering, honed to an ultimate state. The lightweight monocoque blends carbon fiber with aluminum, holding mass to around 1,900 kg for poise and control. Experts suggest the Brouillard could surpass the 282 mph (454 km/h) achieved by the Mistral, aided by more aerodynamic bodywork and the fixed ducktail. Whether or not it sets a new benchmark, its intent is clear: relentless speed paired with refinement. Exclusivity with an exacting timeline The commissioning process is intensive, with roughly 18 months of development leading to a 2027 delivery. Bugatti’s program limits itself to no more than two creations per year, ensuring focus and rarity. Each example leaves the atelier as a singular vision, shaped to the client’s collection and the brand’s ethos. An unveiling is planned for Monterey Car Week 2025, placing the Brouillard amid the world’s most discerning enthusiasts. Its patron is a European collector deeply invested in the Bugatti universe, from historic and modern cars to Carlo Bugatti furniture and Rembrandt Bugatti bronzes. A valuation poised to make history While Bugatti has not disclosed an official price, informed estimates place the Brouillard between €10 million and €30 million. Context matters: a series Mistral begins near €5 million, while the Brouillard adds profound personalization, sculptural changes, and a pace of production that keeps scarcity front and center. In an arena where uniqueness drives value, the path to a brand record is unmistakably open. What it signals for the marque The Brouillard crystallizes Bugatti’s dual identity: traditional savoir‑faire intertwined with boundary-pushing performance. It exhibits a narrative soul—the white horse of Molsheim—woven into modern materials and world-class aerodynamics. As a rolling statement, it reaffirms that the Bugatti idea is larger than speed alone; it is about story, craft, and the courage to pursue an ideal to its logical limit.
Mistral closes in on Big AI rivals with new open-weight frontier and small models
Mistral unveils its Mistral 3 lineup, including a frontier model and efficient small models designed for offline, customizable enterprise use—aiming to prove small, fine-tuned AI can beat closed-source giants.