GLM-4.5-Air

Z.AI introduced GLM-4.5 and GLM-4.5-Air on July 28, 2025, as its latest flagship series targeting reasoning, coding, and agentic workloads. GLM-4.5-Air uses a Mixture-of-Experts design with 106B total and 12B active parameters, and ships with a 128k-token context window and native function calling. Open weights are available on Hugging Face and ModelScope. Both models are hybrid reasoners offering a thinking mode for complex reasoning and tool use, and a non-thinking mode for instant responses, controllable via a reasoning-enabled flag. On Z.AI's internal evaluations across 12 benchmarks covering agentic, reasoning, and coding tasks, GLM-4.5 ranked third and GLM-4.5-Air ranked sixth. The models are accessible via Z.AI's chat interface and Z.AI API.

Z.ai published the official launch blog for the GLM-4.5 model family on 2025-07-28, formally introducing GLM-4.5-Air alongside its larger sibling GLM-4.5. GLM-4.5-Air is described as a 106-billion-total-parameter, 12-billion-active-parameter Mixture-of-Experts model designed to unify reasoning, coding, and agentic work The same blog frames GLM-4.5-Air as the lighter variant of Z.ai's flagship line and reports its overall benchmark standing at 6th place across the 12-benchmark comparison panel covering agentic, reasoning, and coding tests, with agent evaluation done on τ-bench and BFCL-v3. The post is a primary creator announcement ra

Z.ai released GLM-4.5-Air as part of its GLM-4.5 family, a Mixture-of-Experts model designed for autonomous AI agents. The Air variant uses 106 billion total parameters with 12 billion active, activating roughly 11% of parameters per inference for improved efficiency over dense alternatives at similar cost. The model supports hybrid reasoning via a toggleable thinking mode for advanced reasoning and tool use, and a non-thinking mode for real-time interaction. Distribution channels include Hugging Face, OpenRouter, the Z.ai API platform, and chat.z.ai, with Z.ai positioning GLM-4.5 as competitive with proprietary offerings from OpenAI, Anthropic, and Google.

Z.ai (formerly Zhipu) announced GLM-4.5 on July 28, 2025, releasing it as an open-source Mixture-of-Experts family with two variants: the 355B-parameter flagship and the streamlined GLM-4.5-Air with 106B total parameters. The series integrates reasoning, coding, and agentic capabilities into a single model and was evaluated on 12 representative benchmarks, with the flagship ranking third globally and first among open-source models on average. GLM-4.5-Air delivers strong performance among 100B-parameter-scale models, while the series achieves API pricing as low as $0.11 per million input tokens and $0.28 per million output tokens, with a high-speed variant exceeding 100 tokens/sec. The models are designed as Z.ai's first foundation models with native agent capabilities, autonomously planning multi-step tasks and managing end-to-end workflows.

Z.AI's developer documentation describes GLM-4.5-Air as a streamlined Mixture-of-Experts model purpose-built for agent-oriented applications, with 106B total parameters and 12B active parameters per forward pass. It shares a training pipeline with GLM-4.5, starting with 15 trillion tokens of general-domain pretraining followed by code, reasoning, and agent fine-tuning, with reinforcement learning applied for reasoning and coding gains. The model supports a 128K-token context length, up to 96K maximum output tokens, and hybrid reasoning via Thinking and Non-Thinking modes toggled through the thinking.type parameter, with dynamic thinking enabled by default. It is optimized for tool invocation, web browsing, software engineering, and front-end development, and integrates with code-centric agents such as Claude Code and Roo Code through tool invocation APIs.

Unsloth published a Hugging Face repository distributing GGUF quantisations of GLM-4.5-Air under the MIT open-source license, with the page indicating around 15,946 downloads in the preceding month and a 110B-parameter GGUF model size. The repo independently corroborates Z.ai's design claims — 106B total / 12B active M The repository is a third-party quantisation and tooling artifact for GLM-4.5-Air rather than a serving-provider advertisement, and it provides useful, version-specific context: the GLM-4.5 series is presented as a foundation model family for intelligent agents, with GLM-4.5-Air positioned as the more compact member de