AmeforgeAmeforge
Journal
Model ReleaseJune 2026 · 5 min read

Introducing SAM-G

Structured Action Model — Foundation for On-Device AI

SAM-G (Structured Action Model) is the first foundation model released by Ameforge. It was designed from the ground up for one purpose: to produce reliable, structured output on constrained hardware, without any cloud dependency.

Most language models are optimised for open-ended generation. SAM-G is different. It is a dual-mode model that can operate as a conversational assistant or as a structured action planner, switching modes based on a simple prompt token. This design choice keeps the model compact while covering the two core use cases that matter in production robotics and automation: answering questions and emitting executable action plans.

What SAM-G does well

SAM-G was evaluated on ten application domains. It achieves strong results in structured generation, instruction following, and multi-step planning. The model consistently produces valid JSON action plans, maintains coherent state across multi-turn conversations, and handles domain-specific vocabulary in robotics, automation, and industrial control.

Its on-device footprint makes it suitable for deployment on embedded systems, edge servers, and robotic controllers where network access is unavailable or undesirable.

The foundation for Foros and GearCut

SAM-G is the backbone of both Foros and GearCut. Foros uses SAM-G as its action engine, translating natural language instructions into structured robot commands. GearCut uses SAM-G as its reasoning core, interpreting machining parameters and generating precise toolpath configurations.

By sharing a single foundation model across products, we ensure that improvements to SAM-G propagate to every product in the Ameforge ecosystem.

Availability

SAM-G is available on HuggingFace. The model weights, tokenizer, and inference examples are provided. For integration guidance, refer to the research page.