IBM: Granite 4.0 Micro

Generates coherent text responses using a 3B parameter transformer architecture optimized for inference efficiency on resource-constrained environments. The model employs standard causal language modeling with attention mechanisms fine-tuned to handle extended context windows, enabling multi-turn conversations and document-aware responses without requiring GPU acceleration for deployment.

Maintains coherent dialogue across multiple exchanges by processing concatenated conversation history as context in each inference call. The model uses standard transformer attention to track speaker roles, intent shifts, and contextual references across turns, enabling stateless conversation management where the full history is resubmitted with each new user message.

Generates and analyzes code across multiple programming languages by leveraging transformer attention over tokenized source code, with fine-tuning on technical documentation and code repositories. The model can complete code snippets, explain code logic, and generate code from natural language descriptions, using standard causal language modeling without specialized AST parsing or syntax-aware tokenization.

Executes user instructions by conditioning generation on system prompts that define behavior, tone, and task constraints. The model uses standard prompt engineering patterns where system instructions are prepended to user input, allowing dynamic role-playing, task specialization, and output format control through text-based configuration without model fine-tuning.

Provides text generation through OpenRouter's REST API with support for streaming responses via server-sent events (SSE) or polling. Requests are formatted as JSON payloads containing model parameters (temperature, max_tokens, top_p) and conversation history, with responses streamed token-by-token or returned in full, enabling real-time user feedback and progressive output rendering.

Modulates output randomness and diversity through temperature, top_p (nucleus sampling), and top_k parameters passed to the API. Lower temperatures (0.1-0.3) produce deterministic, focused outputs suitable for factual tasks; higher temperatures (0.7-1.0) increase creativity and diversity for generative tasks. The model applies these parameters during token sampling, affecting probability distribution over vocabulary without retraining.

Constrains output length by specifying max_tokens parameter, which limits the number of tokens generated before stopping. The model stops generation when the token limit is reached, even if the response is incomplete, enabling cost control and predictable output sizes. Token counting is approximate (1 token ≈ 4 characters for English text) and handled server-side by OpenRouter.