Tag: foundation models

Prince Mario-Max Schaumburg-Lippe: Aleph Alpha Launches Kolibri Sovereign AI Model

A Hummingbird Lands on German Reunification Day

The timing was deliberate. On October 3, 2026, the Day of German Reunification, Aleph Alpha released Kolibri. Kolibri is German for hummingbird, and the name fits the engineering: a 78.1-billion-parameter model that only ever uses about 3.5 billion of them at a time.

The release landed on the company’s blog under a headline that made no attempt at subtlety: “Kolibri Has Landed: A Sovereign Open-Weight Model.” The full weights went up on Hugging Face the same morning, under the Apache 2.0 license. That last detail matters more than the poetry. Apache 2.0 means anyone can download Kolibri, run it, fine-tune it, and ship commercial products on top of it, without asking permission or filling out a form.

This is Aleph Alpha’s answer to the biggest question in European AI: can Europe build a serious model of its own, on its own terms?

The Engineering in Plain Terms

Kolibri is a mixture-of-experts model. Think of it as a model built from hundreds of specialists instead of one generalist. In each of its 50 layers, a router picks 6 experts out of 384 to handle the current token. Total parameters: 78.1 billion. Active on any given token: roughly 3.46 billion, or about 4.4 percent.

The point of that split is cost. All 78 billion parameters have to live in memory (about 78 GB in FP8, so the minimum hardware is two NVIDIA A100 80GB GPUs), but only a fraction of them burn compute on each token. You get the knowledge capacity of a giant model with the running cost of a much smaller one.

The context window is long: 262,144 tokens natively, extendable to a million through configuration. The model ships with four reasoning effort levels (none, low, medium, high) and tool calling. It handles English and German, with German making up about 21.3 percent of pre-training data, English around 62 percent, and code about 14 percent. Aleph Alpha even built a custom tokenizer, UniBPE, with a 128,000-token vocabulary tuned for German compound words, so German text costs fewer tokens to process.

Training itself is part of the pitch. Kolibri was trained on 768 NVIDIA B200 GPUs in Germany and Finland, under European and German law, on roughly 24 trillion tokens. Before committing to the full run, the team validated the pipeline on a smaller sibling, Kolibri Origin (30 billion total, 3 billion active, 65k context), then scaled the same approach up. Pre-training stayed stable across hardware failures and dropped connections without human intervention, which at this scale is not a small achievement. One faulty node at 768 GPUs usually means a dead run and a 3 a.m. pager.

The Benchmarks, Honestly Framed

Aleph Alpha publishes its numbers, as every lab does, and they should be read the way all vendor benchmarks are read: as a starting point, not a verdict.

Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, and 85.9 on LiveCodeBench v6, per the company’s reporting. The more interesting chart is not a ranking. It plots average score against decoded text per second per GPU, against Kolibri Origin, Qwen3.6-35B-A3B, Nemotron 3 Super, and Mistral Small 4. Aleph Alpha claims Kolibri sits on the Pareto frontier there: best quality for the serving cost, against models with up to four times its active parameter count.

Two details stand out. First, the company ran its math and science benchmarks in German and published that column alongside the English one. Almost nobody does this, and it is exactly what a model pitched at German public administration should be doing. Second, the model is signed to the EU General-Purpose AI code of practice, which is the compliance story European customers actually need to hear.

The model card also notes Aleph Alpha designed Kolibri to refrain from answering when it lacks supporting evidence, an anti-hallucination stance that matters for mission-critical use. Take it as a design goal to verify in practice, not a solved problem.

Sovereignty You Can Download

“Sovereign AI” gets thrown around a lot. Kolibri gives it a concrete meaning. Aleph Alpha uses the word in two senses: how the model was built (by teams in Germany, on infrastructure in Germany and Finland, under European and German law) and how it reaches customers (open weights they can run in their own data centers).

That second part is the real story. It is the same direction the enterprise market has been moving all month. IBM made its coding agent platform self-hostable this week, letting companies keep code and context inside their own walls. Kolibri takes the idea further down the stack: the model itself, downloadable, Apache-licensed, yours to run. No API key. No vendor with a kill switch. No sensitive documents traveling to someone else’s cloud.

“Kolibri demonstrates that we have the talent and expertise in Germany to develop competitive AI models,” said CEO Ilhan Scheer. “For us, AI sovereignty means freedom of choice by retaining the ability to build and advance this technology, and giving customers control over how they use it.”

Why It Matters Beyond Germany

The open-model conversation in 2026 has been dominated by the US and China. Reflection is reportedly about to ship the American answer. DeepSeek and Qwen set the bar the American models are chasing. Kolibri makes the map triangular: a European open-weight model, built under European law, benchmarked in German as well as English, and released under a license that lets businesses actually use it.

For regulated sectors, hospitals, banks, aerospace contractors, and government agencies across Europe, the calculation is simple. The best closed models are brilliant and unusable for your most sensitive data. Kolibri is the attempt to close that gap: frontier-adjacent quality, two GPUs of hardware, your building, your rules.

It will not be the biggest model of 2026. That is not the point. Kolibri is proof that a 200-person team in Heidelberg can ship a serious open-weight model on its own infrastructure, in months rather than years, and hand it to the world under Apache 2.0. The hummingbird landed. Watch what it builds next.

Prince Mario-Max Schaumburg-Lippe: FieldAI Eyes $10B Valuation in $700M Robotics Round

The hottest money in AI right now isn’t going to chatbots. It’s going to robots. FieldAI, the Irvine startup building what it calls a universal general-purpose brain for robots, has signed a term sheet for a $700 million financing round at a $10 billion valuation, according to a Business Insider report published October 2.

Five times. That’s the multiple. FieldAI was worth roughly $2 billion barely a year ago. The new round, which hasn’t formally closed and whose lead investor remains undisclosed, would quintuple that number and put the 2023-founded company in the top tier of private robotics firms, alongside Physical Intelligence at around $11 billion and Skild AI above $14 billion.

What FieldAI actually builds

Here’s the contrarian part: FieldAI makes no physical robots at all. No humanoids, no arms, no wheels. The company sells software, foundation models for robots that let machines navigate and work autonomously in messy, unpredictable environments. One stack powers humanoids, robot dogs, drones, industrial rovers and wheeled vehicles, turning sensor data into continuously updated digital twins of the environment so robots can operate without prior maps, GPS or predefined paths.

That “no maps” detail is the technical pitch. Traditional robot navigation leans on pre-mapped environments, which works fine in a warehouse and falls apart on a construction site where the layout changes daily. FieldAI’s models are designed to account for uncertainty and risk on the fly, adjusting behavior to avoid collisions and navigation mistakes as conditions shift. In March, the company partnered with Boston Dynamics to support the Spot quadruped for industrial inspection tasks, putting its software on one of the most deployed mobile robots in the world.

The customer list is where the story gets its legs. FieldAI says revenue plus signed customer contracts has crossed $135 million across more than 30 customers in construction, data centers, energy and defense, up at least $35 million since June. Construction firms, data center operators and inspection contractors are paying for robot autonomy that works outside the lab. That’s the difference between a demo and a business.

Why investors are paying up

The 5x valuation jump reflects how fast investor appetite has swung from chatbots toward machines that act in the physical world. Robotics startups have drawn a wave of capital this year as foundation models proved good enough to control hardware without constant human oversight, and FieldAI sits at the center of the software layer: the brain, not the body.

CEO Ali Agha brings a resume that helps explain the conviction. He spent seven years at NASA’s Jet Propulsion Laboratory leading autonomy work, including the DARPA Subterranean Challenge, where his team won the urban circuit in 2020. Robots that navigate caves and collapsed tunnels without GPS are a decent audition for robots that navigate construction sites. The company has also been hiring engineers from Google DeepMind, Tesla, Nvidia and Boston Dynamics as the competition widens.

The investor roster doesn’t hurt either. Prior backers include Jeff Bezos’ family office, Laurene Powell Jobs’ Emerson Collective, Khosla Ventures, Nvidia’s NVentures fund and Intel Capital. When that crowd writes follow-on checks, it’s a signal the diligence is real, even if the lead on this round hasn’t been named yet.

The physical AI gold rush

FieldAI’s round is the latest and largest marker in what has become 2026’s defining funding theme: physical AI. The logic runs like this. Language models conquered the digital world; the next frontier is models that operate in the physical one, and whoever owns the software layer under the humanoid and industrial-robot boom owns a platform position.

The numbers tell the story of the frenzy. Robotics trackers have logged over 150 stories in the last 90 days. The comps keep ratcheting upward: NEURA Robotics raised up to $1.4 billion in June at about $7 billion in Europe, Genesis AI was reported raising $500 million at around $3 billion in July, and now FieldAI at $10 billion with a term sheet signed. Whether these valuations reflect fundamentals or FOMO depends on who you ask, but the direction of the money is unmistakable.

There’s a practical side to the boom that gets less attention than the valuations. Training robot brains takes serious GPU capacity and serious data, which is why companies like Sharon AI are borrowing hundreds of millions against their GPUs to build AI factories. And serving the resulting models efficiently is its own industry now, with new inference platforms bringing open-model serving to production scale. FieldAI’s software has to live somewhere, and the infrastructure to run it is being built in parallel.

The honest caveats

Let’s be clear about what’s known and what isn’t. The round hasn’t closed. The lead investor hasn’t been disclosed. The $135 million figure combines recognized revenue with signed contracts, and the report doesn’t break out the split, so treat it as pipeline strength rather than run rate. At $10 billion, FieldAI needs deployments that convert pilots into large recurring contracts. Paper valuations don’t torque motors, as one industry observer memorably put it.

There’s also the integration question. A single software brain that pilots quadrupeds, humanoids, drones and rovers across construction, energy and defense is a massive engineering promise. The environments are different, the sensor suites are different, the failure modes are different. FieldAI’s bet is that foundation-model scale generalizes across all of it. That’s the thesis investors are paying $10 billion for, and it’s still a thesis.

Why this one might be different

What separates FieldAI from most physical-AI pitches is the revenue number, however blended. A lot of robot-brain startups sell a future. FieldAI sells a present: $35 million in new revenue and contracts since June, 30-plus paying customers, a Boston Dynamics partnership, and deployments on real industrial sites. CEO Ali Agha told Business Insider the company has seen “very, very fast growth in the last several months,” and the customer count backs up the claim.

The defense angle deserves a mention too. Construction, energy and defense contractors all show up on FieldAI’s customer list, and dual-use robotics is having a moment as governments look for autonomous systems that work in contested environments. The company doesn’t lead with this, but the investor base, including In-Q-Tel’s peers in the broader ecosystem, suggests it’s part of the thesis.

What to watch next

Three things will tell you whether the $10 billion tag holds. First, who leads the round when it closes, and whether the terms match the reported number. Second, whether FieldAI starts disclosing named customers beyond the anonymized counts, because enterprise logos are the currency of credibility at this scale. Third, the conversion story: pilots to production contracts, contracts to recognized revenue.

The broader trend to watch is the platform battle underneath. FieldAI, Physical Intelligence, Skild AI and a handful of others are all racing to become the operating system layer under the humanoid era. Only one or two will get there, but the winner gets to tax an entire industry’s worth of machines. That’s the $10 billion bet in a sentence.

Robots that work in the real world, not the demo hall, are the whole game. FieldAI just got priced like it’s winning. Now it has to prove it.

Prince Mario-Max Schaumburg-Lippe: General Intuition Raises $220M to Teach AI the Real World

The next frontier of AI isn’t a smarter chatbot. It’s AI that can see, move, and act: pick up a box, navigate a warehouse, climb a set of stairs. And on September 30, one of the biggest bets on that future got a lot bigger: New York–based General Intuition raised $220 million at a $6.2 billion valuation.

The backers are a who’s who of venture capital: Valor Equity Partners, Atreides Management, 776, Point72 Ventures, Khosla Ventures, and General Catalyst. One of the largest physical-AI raises of the month, and a clear signal that serious money is following the agentic-AI wave into the real world.

The capital is earmarked for GPU cluster acquisitions, accelerated foundation-model training, expanded machine learning and reinforcement learning research teams in New York, and commercial infrastructure spanning both virtual gaming systems and physical robotics platforms.

Why Games Are the Gym for Robots

The company’s core idea is elegant. Games were the original training ground for modern AI: think DeepMind learning Atari, AlphaGo conquering Go. General Intuition is betting that millions of hours of gameplay telemetry is the bridge to robots that function in messy reality.

Here’s the clever bit. Instead of hand-labeling the physical world, an expensive, slow, painstaking process, learn intent from players who already demonstrate it. Every game session is a human showing, moment by moment, what they meant to do: navigate this space, grab that object, avoid that obstacle.

General Intuition’s tech is a multi-modal “action foundation model” trained on massive proprietary datasets of multi-angle gameplay video combined with player input telemetry and real-time execution matrices. The model learns spatial navigation, physics interactions, and operational intent, the same skills a robot needs, and drives both autonomous agents in simulations and humanoid robots in the real world.

In other words: the training data is hiding in play. Humans already generate exquisitely detailed demonstrations of physical intent every time they game. General Intuition is just harvesting it.

The Embodied AI Wave Is Building

This raise doesn’t exist in isolation. Physical AI, the industry term for AI that acts in the physical world, is having its moment. Humanoid robots like Agility’s Digit are getting stronger, safer, and more capable. Driverless trucks are on public roads. Robotaxi fleets are scaling fast.

Each of those machines needs a brain that understands physics, not just language. A chatbot can be wrong and it’s a joke; a 200-pound humanoid can be wrong and it’s a lawsuit. The bar for “good enough” in physical AI is brutally higher than in text, which is why the training approach matters so much.

General Intuition’s angle, learning from demonstrated intent at massive scale, sidesteps the biggest bottleneck in robotics: labeled real-world data is scarce and expensive. Gameplay telemetry is abundant and rich. If the transfer from virtual to physical works, it’s a shortcut around years of slow data collection.

What $6.2 Billion Says About the Moment

Valuations this size say investors believe embodied AI is following the same arc as language AI: a period of expensive foundational work, then a sudden unlock. The GPU clusters, the expanded research teams, the commercial infrastructure across gaming and robotics: this is a company building the full stack, not a demo.

The New York angle is nice too. The company is expanding its ML and reinforcement learning research teams in the city, planting a flag for physical AI on the East Coast in an industry that tends to default to the Bay Area.

Where the Robots Go First

The commercial infrastructure piece of the raise deserves attention. General Intuition isn’t just training models — it’s building the deployment pipeline across virtual gaming systems and physical robotics platforms. The near-term beachhead is likely the warehouse: structured enough to be tractable, labor-hungry enough to pay for automation. Longer term, the same action models that navigate a game level could navigate a disaster site or a factory floor.

That’s the bet the investors are making with $220 million: that “action” becomes a foundation-model category the way language did, and that the company holding the best action model holds a position worth far more than $6.2 billion. It’s early. But every major AI platform started with someone training an expensive model on data nobody else thought to collect.

The Optimist’s View

Picture where this leads. Robots that learn movement the way humans do, by watching and doing at scale, could take on the dull, dirty, and dangerous work that’s hard to staff: warehouse shifts, disaster cleanup, elder care assistance. The path from a game controller to a helpful humanoid is longer than a press release makes it sound, but the direction is right, and $220 million is a serious down payment.

General Intuition’s bet is simple and, in retrospect, may look obvious: the internet taught AI to think; play will teach it to move. The funding announced today suggests a lot of very smart investors agree.