Z.ai: GLM 4.5 Air

GLM-4.5-Air processes multi-turn conversations with native support for structured function calling via schema-based tool definitions. The model uses a Mixture-of-Experts (MoE) architecture where only a subset of expert parameters activate per token, reducing inference latency while maintaining reasoning quality. It routes conversation context through sparse expert layers, enabling efficient handling of tool invocations, parameter extraction, and agent decision-making without full model activation.

GLM-4.5-Air handles extended conversation histories through optimized token management and sparse attention patterns enabled by its MoE architecture. The model compresses context representation by routing only relevant context through active experts, reducing the computational cost of maintaining long conversation state. This allows multi-turn dialogues with hundreds of messages without proportional latency degradation.

GLM-4.5-Air can generate responses conforming to strict JSON schemas or structured formats through constrained decoding and schema-aware token routing. The model uses its MoE architecture to specialize certain experts for structured output generation, ensuring responses match predefined schemas without post-processing validation. This enables reliable extraction of entities, relationships, and structured information from unstructured text inputs.

GLM-4.5-Air supports server-sent events (SSE) streaming where tokens are emitted as they are generated, enabling real-time response display and token-level monitoring. The model streams through its MoE layers, allowing clients to observe token generation in real-time and implement early-stopping logic based on partial outputs. This architecture enables interactive applications where users see responses appearing incrementally rather than waiting for full generation.

GLM-4.5-Air maintains multilingual reasoning capabilities through language-specific expert routing in its MoE architecture. The model activates different expert subsets depending on input language, enabling code generation, mathematical reasoning, and logical inference across programming languages, natural languages, and formal notations. This approach avoids the parameter bloat of dense multilingual models by specializing experts per language family.

GLM-4.5-Air's MoE architecture dynamically activates only a subset of expert parameters per token, reducing computational cost compared to dense models. The model routes each token through a gating network that selects 2-4 active experts from a larger pool (typically 64-128 experts), achieving inference cost reduction while maintaining output quality. This sparse activation pattern is transparent to users but directly impacts per-token pricing and latency.