somnia-agents-llm-inference

Deterministic LLM inference on-chain uses Qwen3-30B with fixed seed and temperature=0 to guarantee byte-identical outputs across validators, allowing smart contracts to safely branch on AI decisions without breaking consensus.

You can use the inferString function with allowed-values constraints to force LLM outputs into a closed set, ensuring safe and deterministic branching logic within your smart contracts.

Smart contracts use the on-chain tool yield/resume mechanism where the contract yields calldata back to callers for tool execution, then resumes the multi-turn conversation by correctly handling the finishReason and tool_calls resume state.

Yes, the inferToolsChat function supports MCP tool calling for deterministic AI outputs, enabling auto-executed MCP tool calls for real-world data or deterministic tool yields within multi-turn chat conversations.

Yes, on-chain LLM inference supports both string classification and bounded integer outputs through typed inference APIs, specifically inferString for classification and inferNumber for bounded integer inference.

Limitations of using deterministic LLM inference include the strict requirement for careful prompt determinism and correct handling of finishReason and tool_calls resume state, as deviations can break byte-identical consensus across validators.