A 27B powerhouse distilled from Claude 4.6 Opus. Built for deep analytical reasoning and stable agentic performance in the watch market.

A highly efficient 4B parameter model from NVIDIA, optimized for low-latency on-device tasks and high-quality text generation.

A 2026-native reasoning model distilled from R1. Specialized for agentic "Chain of Thought" logic on local hardware.

The mid-2026 flagship using Engram memory architecture, specializing in 1M+ token code generation and autonomous refactoring.

Task-specialized 4B model for natural-language-to-SQL conversion. Distilled from DeepSeek-V3. Quantized GGUF for local database agents.

Google DeepMind multimodal instruction model. 4.5B effective params, 128K context, text+image+audio. Native function calling, configurable thinking modes, Apache 2.0.

OpenClaw-recommended general-purpose model. 7B params, 128K context, MIT license. Balanced speed/quality for daily assistant tasks, research, and multi-step reasoning.

OpenClaw-compatible open-weight 20B model. 64K+ context, Apache 2.0. Balanced performance for tool-use, memory persistence, and multi-channel agentic workflows.

Liquid AI hybrid architecture: 10 LIV convolution blocks + 6 GQA layers. 350M params, 32K context, 9 languages. 313 tok/s CPU decode, <1GB RAM. Day-one llama.cpp/MLX/vLLM support.

Liquid AI hybrid architecture via Unsloth. 700M params, 32K context, CPU-optimized. Built for narrow-scope agentic tasks: data extraction, RAG, multi-turn workflows.

A compact chat-tuned model, distilled from LLaMA-2 for quick interactions.

Mistral AI edge-optimized 3.4B+0.4B vision model. Native function calling, JSON outputs, 256K context. Built for tool-using agentic pipelines.