Models & robotics

Models & robotics

The models themselves and the machines they drive: open-weight and small language models, OCR, and robotics foundation models for embodied AI.

Open-weight models

LLM families released with downloadable weights you can run, fine-tune, and self-host.

• DeepSeek - Lab releasing reasoning- and coding-focused models with weights published under the MIT License.
• Gemma - Google DeepMind's family of lightweight open models built from the same research as Gemini, released under a permissive license.
• GLM - Open-weight General Language Model family from Z.ai, formerly Zhipu AI, targeting long-horizon agentic and coding tasks.
• gpt-oss - OpenAI's open-weight reasoning models released under Apache 2.0 and designed to run on single-GPU or consumer hardware.
• Kimi - Moonshot AI's Kimi family of large mixture-of-experts open-weight models, strong on agentic and coding tasks.
• Llama - Meta's widely deployed open model family, with natively multimodal mixture-of-experts architectures.
• MiniMax - Lab releasing frontier-class open-weight reasoning and coding models under MIT-style licenses.
• Mistral - Lab releasing open-weight dense and mixture-of-experts models under Apache 2.0.
• Nemotron - NVIDIA's open model family released with open weights, training data, and recipes for agentic reasoning.
• Qwen - Alibaba's open-weight family, one of the most-downloaded open model lines, spanning phone-sized to very large MoE models.
• Whisper - OpenAI's open-source automatic speech recognition model for multilingual transcription and translation.

Small language models

Compact models tuned to run efficiently on-device or on modest hardware.

• AlphaEdge - Sovereign-AI startup that compresses and distills models into small, efficient ones for on-prem document and enterprise workflows.
• Phi - Microsoft's family of small open-weight models emphasizing strong reasoning at small parameter counts.
• Pleias - Lab training small open-weight models exclusively on public-domain and permissibly licensed data, alongside the fully open Common Corpus dataset.
• SmolLM - Hugging Face's fully open family of small text and vision models, with training data and code released for on-device use.

Decision models

Non-autoregressive models that skip text generation and return typed, calibrated decisions instead, for routing, classification, and other structured judgment calls.

• Jev - TypeSafe AI's System One model that returns typed, calibrated decisions instead of text, for automated routing, classification, and safety checks.
• Laya - Open-weight, Apache-licensed decision model that returns typed choices, scores, and yes/no judgments with calibrated probabilities across 100+ languages.

Robotics and embodied AI

Foundation models and companies bringing learned control to humanoids and real-world machines.

• 1X - Humanoid robotics company building the Neo home robot around neural control policies.
• AMI Labs - Yann LeCun's lab building world models that learn abstract representations of the physical world so agents can predict and plan their actions.
• Figure - Humanoid robotics company training general-purpose manipulation models for commercial deployment.
• Genesis AI - Physical-AI lab building robotics foundation models and a general-purpose robot that reasons, plans, and acts in the real world.
• Gobano Robotics - Robotics company applying imitation learning, reinforcement learning, and world models to dexterous industrial and logistics tasks like folding, assembling, and bin picking.
• Physical Intelligence - Research lab building foundation models for robot control across embodiments.
• Skild AI - Developer of a general-purpose robot brain trained across thousands of hours of robot data.
• UMA - Physical-AI company building humanoid and mobile robots that learn new skills from demonstration.
• Wandercraft - Robotics company building AI-powered self-balancing exoskeletons and the Calvin humanoid robot for industrial tasks.

OCR

Models that convert documents and images into structured, machine-readable text.

• DeepSeek-OCR - Open-weight OCR model using optical token compression for high-throughput, low-cost document-to-Markdown conversion; runs on vLLM.
• Mistral OCR - Hosted OCR API with strong accuracy on complex tables and handwriting for managed, no-ops document extraction.