AI

Robots Learn to Pack Cars and Read DNA as Open Weights Keep Flowing

By Crypt0 · Nov 11, 2026

Tuesday and Wednesday turned the frontier physical. Google DeepMind shipped a robot brain that packs car parts on a factory floor and a gene model that reads DNA cell by cell, Runway handed the open source world a robot model trained on video, and Thinking Machines shipped a coding model that trains other models by itself. The thread tying the day together is simple. The frontier moved from the chat window into the body, the factory, and the lab in a single day, and the most useful models now arrive as weights you can download and run.

Gemini Robotics 2.5 leads the pack. DeepMind debuted the model on November 10, calling it the most capable vision language action model to date for building a robot brain, and it runs on Gemini 4 Argon, the frontier model that rolled out in late September. The model was tested across more than 200 distinct tasks, completing them 20 percent to 30 percent faster than the previous version, with success rates from 30 percent to 100 percent depending on the task. In a real world proof of concept at Mercedes Benz, Apptronik Apollo 2 humanoid robots packed automotive parts with a 96 percent success rate. The model rolls out through robotics partners Agile Robots, Apptronik, and Boston Dynamics, and it works across many robot types, from parallel grippers to humanoid hands.

AlphaGenome 2 brings the same precision to biology. DeepMind announced the model on November 11, and it predicts how DNA variants affect gene regulation in individual cell types and states, trained on more than 100 million individual cell profiles. On TenK10K, a single cell immune dataset spanning 28 cell types, the model recovered 30 percent of putatively causal expression linked variants at 90 percent precision, compared with 16 percent for AlphaGenome 1 fine tuned on single cell data. The model also predicts DNA methylation, a core piece of epigenetics, and it adapts directly to new single cell data, predicting how DNA variants act in newly profiled cell states even from very small samples. It is available through the AlphaGenome API for noncommercial research, with commercial research in private preview, and it stays a research tool, with the model unapproved for clinical or diagnostic use.

Runway Praxis 1 turns video into robot motion. Runway released the model on November 11 with open weights under the Apache 2.0 license, including for commercial use, and it is a 27 billion parameter mixture of experts diffusion transformer that predicts video and robot actions together. On the Nvidia RoboLab 120 leaderboard, which covers 1,200 test episodes, Praxis 1 reached 46.2 percent success, the top result among world action models and open source robot foundation models. The release includes weights plus inference and fine tuning code on GitHub and Hugging Face, with native LeRobot support and a vLLM Omni serving option. It runs on Nvidia hardware from H100 and B200 data center chips to RTX 5090 workstation cards and Jetson Thor inside the robot itself, and three launch partners, Noble Machines, Standard Bots, and Ultra, already run it on a legged humanoid, a six axis industrial arm, and a dual arm warehouse robot.

Thinking Machines Inkling 2 Small closes the loop on agents that build. The lab introduced its second model generation on November 11, a mixture of experts model with 594 billion total parameters and 17 billion active, supporting image inputs and variable effort. It scored 31 on the Artificial Analysis Intelligence Index, with strong gains in agentic coding, vision, knowledge work, and reasoning. It is available through Tinker, with closed weights at first and full weights under Apache 2.0 next week after external safety testing and red teaming. In one demonstration, the model autonomously trained a Qwen3.5 4B model on Tinker to predict time series like weather and Wikipedia page views, working for 4.5 hours on $426 of Tinker spend and more than doubling the forecasting score. DeepMind AlphaProof Code adds the proof layer, using Gemini powered agents and Lean to formally check properties of Signal deployed cryptographic implementation in four weeks. Meituan added its own open weight giant, LongCat 2.5, a 1.6 trillion parameter mixture of experts model with about 48 billion active parameters per token under an MIT license.

For readers, the open weight wave is now a physical wave. Robot brains, gene models, and coding agents ship as downloadable weights that run on hardware you can buy, and the labs keep safety testing visible while they ship. The practical move is to watch what builders do with Praxis 1 and Inkling 2 Small in the next few weeks, because open weights turn a launch day into a building season, and the builders are already moving.

Sources

• SiliconANGLE DeepMind Gemini Robotics 2.5 and AlphaGenome 2
• Runway Praxis 1 now available open weight robotics model
• HPCWire Thinking Machines Lab introduces Inkling 2 Small
• Unite AI DeepMind AlphaGenome 2 cell state variant effects

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.