axiom-foundation-models-ref

What it does

Describes a Foundation Models framework for on-device large language models with a 3B parameter, 2-bit quantized model and 4096 token context. It covers the LanguageModelSession core class, the @Generable macro for structured output, @Guide constraints, streaming via PartiallyGenerated types, and a Tool protocol to integrate external data fetches. It includes code examples for creating sessions (with or without tools or specific models), using respond(to:) and respond(to:generating:), transcript maintenance, isResponding gating, and streaming integration with SwiftUI. It also explains tool usage patterns (GetWeatherTool example), how tools are attached to a session, and how tool results are incorporated into model responses. It lists navigation help, isResponding behavior, and notes about multiple tools and parallel calls, along with best practices for tool naming and usage.

How it works

The skill defines a Foundation Models framework API reference that includes:

• Model specs: 3B parameter model, 2-bit quantized, 4096 token context; runs on-device with no network or data leaving device.
• LanguageModelSession: maintains transcript, supports multi-turn interactions, allows creating sessions with instructions, tools, or a specific model/use case.
• respond(to:) and respond(to:generating:): functions to generate text or structured outputs, returning a Response with a .content property.
• Generable macro: enables structured outputs via Swift types, with constrained decoding to ensure schema adherence; supports primitive, collection, and composed types; demonstrates usage on structs and enums.
• Streaming: uses snapshot streaming via Itinerary.PartiallyGenerated to progressively fill properties; includes streamResponse and SwiftUI integration.
• Tool Protocol: defines a Tool protocol to fetch external data, with a GetWeatherTool example; describes how tools are attached to a session and how tool outputs are inserted into transcripts.
• Tool calling behavior: tools can be called multiple times per request, possibly in parallel; model decides when to call; arguments are ensured valid via @Generable.

When to use it

Use this reference when you are implementing Foundation Models features, understanding API capabilities, looking up specific code examples, planning architecture with Foundation Models, migrating from prototype to production, or debugging implementation issues. Related skills include axiom-foundation-models and axiom-foundation-models-diag.

What it can touch

Tools are attached to a LanguageModelSession (e.g., GetWeatherTool) to fetch external data. The protocol defines a call(arguments:) async throws -> ToolOutput method; tool outputs can be natural language or structured GeneratedContent. The example shows attaching GetWeatherTool to a session and issuing a prompt to retrieve weather data.

Caveats

The reference notes the model is a 3B parameter, 2-bit quantized, with a 4096 token context; it runs entirely on-device with no network, no cost, no data leaves the device. It emphasizes structured decoding to guarantee correctness and warns about token counts affecting latency via longer prompts. It does not provide outcomes beyond the described behaviors and examples.