@automatey-org/nemotron

Nemotron / vLLM provider helpers for the Vercel AI SDK — the one place that knows the Nemotron chat template (thinking budget) and how to infer a model's modalities. Shared by the CLI, the web app, the extension, and the Unreal toolkit.

import { buildNemotronModel, inferNemotronCapabilities } from '@automatey-org/nemotron';

const model = buildNemotronModel({
  baseUrl: 'http://localhost:8002/v1',
  model: 'nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4',
  apiKey: 'not-needed',
  // enableThinking: true,   // default
  // thinkingBudget: 4096,   // default
});

inferNemotronCapabilities('…-Omni-…-Reasoning-…'); // { vision: true, reasoning: true, omni: true }

• buildNemotronModel(config) returns the concrete model (ready for wrapLanguageModel / @automatey-org/agent-core), injecting chat_template_kwargs { enable_thinking, thinking_budget } via transformRequestBody.
• inferNemotronCapabilities(modelId) → { vision, reasoning, omni }.

Peer dependency: @ai-sdk/openai-compatible >= 3.