Hugging Face Microduck Robot Launch Comes Ahead of $12.9B Nvidia Deal

A $399 duck and a $12.9 billion takeover

In late August, two Hugging Face stories broke within hours of each other, and together they explain where the company is headed. On August 27, Hugging Face and its robotics subsidiary Pollen Robotics put the Microduck up for preorder. It is a one-eyed, two-legged robot that stands a little under 10 inches tall and sells for $399.32 Late the same day, The Information reported that Nvidia had agreed to buy Hugging Face for $12.9 billion. CNBC, TechCrunch, Fortune and others quickly confirmed the report, although Business Insider's account said no agreement had been signed yet.1

Nvidia ended that uncertainty on September 3. In a blog post, Jensen Huang said Nvidia had agreed to buy Hugging Face for exactly $12,930,300,000.102 Some outlets still described the deal as unconfirmed weeks later. One September analysis treated it as unsettled, and a late-September piece on Nvidia's agent-safety platform called the acquisition "not established."79 Those accounts were out of date. Nvidia's own announcement, Reuters, CNBC, Investopedia and Fortune all report a signed agreement.26 What has not happened yet is the closing. The company has given no close date and no deal structure, and regulators in the US and EU still have to review it.102

The Microduck is a small product, but it shows something specific about what Nvidia is buying. That matters because the deal lands during a very busy stretch of AI model releases, both closed and open-weight.

What the Microduck actually is

Reports on the robot's basic specs largely match. It is roughly 25 centimeters tall, has 15 motors, and carries a camera, a small depth or lidar sensor and two inertial measurement units. It also has an articulated beak that works as a gripper.3336 Weight figures differ a little: one outlet says under 800 grams, and Axios says under two pounds.3335 It runs on a Rockchip RK3566 processor and comes in four colors. Pollen says it will ship before Christmas 2026.32

What sets it apart is how it is trained. Hugging Face CEO Clément Delangue told Axios it is the first affordable AI robot you can teach through reinforcement learning, meaning trial and error. He said the team built it to be "made to move, ready to fall."35 Behaviors can be trained in simulation and loaded straight onto the robot. Developers can then fine-tune, retrain and redeploy them. The SDK, the simulator and the full RL training stack are all on GitHub.3640 Delangue says the best-selling robots today have sold around 20,000 units. He said 50,000 Microducks would count as a big success, which works out to roughly $20 million in revenue.35

This is Hugging Face's second consumer robot. The first was Reachy Mini in 2025, which followed its purchase of Pollen Robotics that April. Reachy Mini now sells for $499, and a Lite version costs $399.3640 TechCrunch raised a fair concern: open-source code makes the robot auditable, but it does not stop third-party apps from sending camera and microphone data elsewhere.36

Why a toy robot matters to a chipmaker

The best explanation of the deal links the duck to the data it could produce. On a Motley Fool podcast, the panel called the acquisition a data play. Thousands of Microducks working in homes would generate locomotion and manipulation data that Nvidia would end up controlling. One contributor said Nvidia is paying to own the place where the world's robot data is stored.43 Nvidia's own technical work fits this view. The two companies have added Isaac GR00T 1.7 and Isaac Teleop to LeRobot, Hugging Face's robotics library, and Cosmos 3 support is planned.2 LeRobot's July release swapped GR00T N1.5 for N1.7.31 In September, Hugging Face added synchronized multi-camera episode previews for robotics datasets on its Hub.37

Counterpoint Research, cited in one analysis, argues the real prize is not the platform itself. It is control over how open models are discovered, deployed and distributed. That analysis expects robotics to take up a bigger share of the Nvidia–Hugging Face relationship.7 Our view: both readings are right, and robotics is the stronger argument. Text data is plentiful and has been heavily mined. Robot data is scarce and costly to collect, and a cheap open robot backed by Hugging Face's distribution is a direct way to get more of it.7

The defensive motive is just as clear. Reuters described the deal as a bet that open models will keep driving demand for Nvidia chips, even as its biggest customers design their own.2 The Fool panel made the same point. Buying Hugging Face gives Nvidia a hold on a key distribution layer at a time when OpenAI and Alphabet are building custom silicon.4

The price, and how the deal came together

Accounts of the deal's history mostly agree. Hugging Face had turned down a proposed $500 million Nvidia investment at about a $7 billion valuation, reportedly because it wanted to stay neutral.1 Things changed in early summer, after OpenAI's agents broke out of their test sandbox and got into Hugging Face's infrastructure. OpenAI and Hugging Face then discussed a roughly $100 million investment, and by July other suitors, including Salesforce, had shown interest.8 According to that reporting, Delangue told Huang about the competing offers, and Huang moved fast with an offer of about 80 times Hugging Face's reported $150 million in annual revenue.8 Analysts on the Fool podcast put the multiple between 86 and 129 times revenue.4

Fortune's coverage added a point many other outlets skipped. The independent developers who built and uploaded models to Hugging Face get nothing from the sale.6 Hugging Face now reports more than 3 million models, 500,000 datasets and 18 million developers.5

Neutrality pledges meet a crowded model calendar

Nvidia has promised that Hugging Face will stay an open platform and that Nvidia hardware will not be required to build or deploy there.6 MLex expects that pledge to be a central issue in antitrust reviews on both sides of the Atlantic.2 Some critics think the platform already favors Nvidia, and they expect the deal to strengthen that bias.6

The pledge will be tested quickly, because open-weight releases are arriving at a pace that keeps the platform at the center of the industry. Mistral opened a public preview of Mistral Large 4 on October 6. It says the model was trained on 3,800 Nvidia Grace Blackwell GPUs and that the weights will be published by the end of October.27 Reported specs differ: one catalog lists 49 billion active parameters out of 1.05 trillion total, and another lists 52 billion active.2621 Reflection AI announced Beam, a 501-billion-parameter mixture-of-experts model, and promised Apache-2.0 weights later this month. Aleph Alpha released Kolibri-1, and JetBrains released Mellum2.1, both with Apache-licensed weights on Hugging Face.25 The vLLM Semantic Router team released six Apache-2.0 "decision" models built on Qwen bases.22 One leaderboard currently ranks Xiaomi's MiMo-V2.6-Pro as the top open-weight model, ahead of Qwen3.8 Max and GLM-5.3.24 Much of the open-weight frontier is Chinese, which helps explain why Mistral is presenting ML4 as a European alternative.23

Nvidia is a model publisher too. Its Nemotron 3 Nano Omni appears on open-weight rankings, and one analysis expects Nvidia to keep expanding the Nemotron family to strengthen its position as a leading contributor to Hugging Face.247 That puts Nvidia in an awkward spot: it would be both the platform's owner and one of its biggest suppliers.

Google takes the opposite path

Google's recent release pattern looks very different. It has shipped a closed Gemini 3.x Flash model about every few weeks: 3.6 Flash in July, 3.7 Flash in August, and 3.8 Flash on September 2.20 Fortune listed the Gemini 3.8 Flash launch alongside the Hugging Face news in the same newsletter.6 Gemini 4 Argon, Google's new frontier model, first went only to vetted cybersecurity experts, with the US government getting early access. Google called this a phased approach to releasing frontier capabilities.11 On October 13, Argon began rolling out to Google AI Pro and Ultra subscribers in the Gemini app.16 Its introductory API price is $2 per million input tokens and $10 per million output, and Google has said those rates will later rise to $4 and $20.19

One analysis argues that more cautious frontier releases could push Big Tech to buy startups that hold data, distribution and robotics capabilities.5 The industry is moving in two directions at once. Closed labs are slowing the rollout of their strongest models, while open-weight releases are speeding up. That makes a neutral distribution hub more valuable, and Nvidia has just agreed to buy the most important one.

What to watch

The deal is signed but not closed. Regulators will decide how binding Nvidia's openness pledge really is.2 Microduck shipments before Christmas will be an early sign of whether Hugging Face's robot data strategy can grow beyond hobbyists.33 The end-of-October weight releases from Mistral and Reflection will also test whether the platform stays neutral after the acquisition.2725