Z.ai's GLM-4.5-Air Packs 106B Agent Reasoning Into 12B Active Compute
Z.ai's 106B Mixture-of-Experts model activates just 12B parameters per token, trades a sliver of depth for cheaper inference, and ships under MIT.
- GLM-4.5-Air is a 106B MoE model with 12B active parameters, released by Z.ai under MIT license
- Hybrid reasoning: thinking mode for agents and tool use, non-thinking mode for fast chat
- 128K context, 96K max output, trained on 23T tokens with expert iteration and RL curriculum
- Scores 59.8 average across 12 benchmarks, 57.6 on SWE-Bench Verified, 81.4 on MMLU-Pro
- Priced at $0.20/$1.10 per million tokens on Z.ai API, available at 43 providers
- Technical report and weights on Hugging Face, with vLLM and SGLang support
At publication, Z.ai’s model repository showed nearly 800,000 downloads and more than 3,900 likes. GLM-4.5-Air packages the company’s reasoning, coding, and tool-use capabilities into a smaller sibling of the flagship GLM-4.5.
Developers get downloadable weights under the MIT license, a 128K-token context window, and two inference modes in one checkpoint. The main tradeoff is memory: sparse routing reduces computation per token, but serving still requires access to all 106 billion parameters.
Sparse routing does the heavy lifting
GLM-4.5-Air is a mixture-of-experts transformer with 106 billion total parameters and about 12 billion active for each token. Its router sends each token through a small subset of feed-forward experts, reducing arithmetic compared with a dense 106B model.
The 12B active figure describes computation rather than storage or exact dense-model performance. Every expert must remain available in memory, and multi-GPU communication, attention layers, routing, and the key-value cache all contribute to latency.
| Component | Specification |
|---|---|
| Total parameters | 106 billion |
| Active parameters | About 12 billion per token |
| Transformer layers | 46 |
| Expert routing | Eight routed experts plus one shared expert per token |
| Attention | 96 query heads and eight key/value heads |
| Hidden size | 4,096 |
| Vocabulary | About 151,600 tokens |
| Activations and normalization | SwiGLU and pre-normalization with RMSNorm |
| Context window | 128K tokens |
| Maximum advertised output | 96K tokens, subject to the runtime’s sequence limit |
| License | MIT |
Two reasoning paths, one checkpoint
Thinking mode allocates additional generation to multi-step reasoning, coding, and tool selection. Non-thinking mode produces direct responses for latency-sensitive chat, extraction, classification, and other tasks that rarely benefit from a long reasoning trace.
Applications can switch modes per request without loading another model. In the Hugging Face chat template, the enable_thinking argument controls the behavior:
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True,
)This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
