Google: Gemma 4 31B

Processes both text and image inputs simultaneously within a single inference pass, using a unified embedding space that aligns visual and textual representations. The model architecture integrates a vision encoder (likely ViT-based) with the language model backbone, allowing it to reason across modalities without separate encoding steps. Supports up to 256K token context window for extended reasoning over mixed-media documents.

Implements a two-stage inference architecture where an optional 'thinking' mode enables the model to perform internal chain-of-thought reasoning before generating final outputs. When activated, the model allocates computational budget to explore solution spaces, backtrack, and refine reasoning before committing to a response. This is configurable per-request, allowing callers to trade latency for reasoning depth on complex problems.

Implements OpenAI-compatible function calling interface where the model can request execution of external tools by generating structured function calls based on a provided schema registry. The model learns to map natural language intents to function signatures, parameter types, and argument values during training. Supports multiple concurrent function calls per response and integrates with standard tool-use patterns (function name, arguments object, return value handling).

A 30.7 billion parameter dense transformer model optimized for efficient inference on commodity hardware and cloud accelerators. The 256K token context window is achieved through efficient attention mechanisms (likely grouped query attention or similar) that reduce memory overhead while maintaining full context awareness. The dense architecture (no mixture-of-experts) ensures predictable latency and memory usage without routing overhead.

The 'IT' (Instruction-Tuned) variant is fine-tuned on instruction-following datasets and RLHF (reinforcement learning from human feedback) to produce helpful, harmless, and honest responses. The model learns to refuse harmful requests, acknowledge uncertainty, and provide structured outputs when appropriate. Safety training is integrated into the model weights rather than applied as a post-hoc filter, enabling more nuanced safety decisions.

Supports efficient batch processing of multiple requests with different input lengths through dynamic padding and attention masking. The model can process heterogeneous batch sizes (e.g., 5 short queries and 3 long documents in the same batch) without padding all inputs to the longest sequence length. This is achieved through efficient attention implementations that skip padding tokens and optimize memory layout.

The model can be constrained to generate outputs matching a provided JSON schema, ensuring structured data extraction without post-processing. This is implemented through constrained decoding where the model's token generation is restricted to valid continuations that maintain schema compliance. The model learns during training to map natural language to structured outputs, and inference-time constraints prevent invalid JSON or schema violations.