GLM-4.5-Air: MIT Licence, 128K Context and Fine-Tuning Hardware

What GLM-4.5-Air is

GLM-4.5-Air is the smaller member of Z.ai's GLM-4.5 release, published under the MIT licence with commercial use and secondary development explicitly permitted, and carrying a 128K token context window. Like its larger sibling it ships as a base model, a hybrid reasoning model and an FP8 build of the reasoning model — a completeness most open-weight releases do not match. At roughly $0.20 per million input tokens and $1.10 output through third-party providers it costs about a fifth of GLM-4.5 to rent.

Fine-tuning hardware

Z.ai publish what it takes to fine-tune both models. GLM-4.5 requires 16 H100 GPUs for LoRA fine-tuning at a batch size of one per GPU. GLM-4.5-Air requires 4.

Benchmarks

Z.ai's own aggregate across twelve industry-standard benchmarks puts GLM-4.5 at 63.2 and GLM-4.5-Air at 59.8, which they describe as competitive results with superior efficiency. On LMArena's text leaderboard GLM 4.5 Air has 1373 Elo (rank 199); by category, 1426 Elo in Coding (rank 194), 1386 Elo in Math (rank 187), 1390 Elo in Expert (rank 197) and 1327 Elo in Creative Writing (rank 208). GLM-4.5 has 1411 Elo (rank 153).

Architecture and modes

Z.ai's card gives GLM-4.5-Air 106 billion total parameters with 12 billion active per token, against GLM-4.5's 355 billion and 32 billion. Both are hybrid reasoning models with two modes: a thinking mode for complex reasoning and tool use, and a non-thinking mode for immediate responses. When served through vLLM or SGLang, thinking is enabled by default and is switched off by passing enable_thinking set to false in the chat template arguments. Both support tool calling in the OpenAI-style tool description format. The model code, tool parser and reasoning parser are implemented in Transformers, vLLM and SGLang, and the weights are on Hugging Face and ModelScope in BF16 and FP8 builds, with a separate Base checkpoint.

Hardware for inference

Z.ai publish minimum configurations for full-featured inference with SGLang: GLM-4.5-Air runs on four H100s or two H200s in BF16, and on two H100s or one H200 in FP8. Using the full 128K context needs eight H100s or four H200s in BF16, or four H100s or two H200s in FP8. These figures assume multi-token prediction layers with speculative decoding enabled, no CPU offload, a batch size of at most eight, hardware with native FP8 support, and server memory above 1TB. With the Swift framework on 96GB H20 GPUs, Z.ai list four GPUs for LoRA and 32 for full supervised fine-tuning or reinforcement learning.

On the Z.ai API

Z.ai's API documentation gives the GLM-4.5 series a 128K context length and up to 96K output tokens, with text input and output. It lists thinking mode toggled through the thinking.type parameter, with dynamic thinking enabled by default, along with function calling, streaming, context caching and JSON structured output. Z.ai state the series was pre-trained on 15 trillion tokens of general-domain data, then fine-tuned on code, reasoning and agent tasks and trained with reinforcement learning. Besides GLM-4.5 and GLM-4.5-Air, the API offers GLM-4.5-X and GLM-4.5-AirX as faster variants and GLM-4.5-Flash as a free tier.

Official Resources