Model
Model explorer

GLM-4.5-Air (106B-A12B)

OPEN
Zhipu AI / Z.ai (Tsinghua) · GLM-4.5 family · released Jul 28, 2025

Compact MoE sibling of GLM-4.5 for efficient agentic deployment; scored 59.8 vs GLM-4.5's 63.2 on Zhipu's 12-benchmark composite at launch.

ReasoningCodingVisionFunction callingTool useAgentic
1617.7
Elo · rank #99
Parameters
106B
Active params
12B (MoE)
Context
128K tokens
Architecture
Mixture-of-Experts, 106B total / 12B active, hybrid thinking/non-thinking response modes
License
MIT
Languages
—
API price (in/out)
$0.2 / $1.1
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath89.4%#47
best: GPT-5.2 · 100.0%
BFCL v3Agents76.4%#3
best: Hunyuan-A13B · 78.3%
BrowseCompAgents21.3%#41
best: Kimi K3 · 91.2%
GPQA DiamondReasoning75.0%#104
best: GPT-5.6 · 94.6%
Humanity's Last ExamReasoning10.6%#87
best: Claude Sonnet 5 · 57.4%
IFBenchReasoning37.6%#56
best: MiniMax M3 · 83.0%
IFEvalReasoning86.3%#47
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding70.7%#67
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math98.1%#5
best: GPT-5 · 99.4%
MMLU-ProKnowledge81.4%#50
best: Claude Fable 5 · 91.5%
MMLUKnowledge87.4%#32
best: OpenAI o3 · 92.9%
SimpleQAKnowledge14.5%#27
best: GPT-4.5 · 62.5%
SWE-bench VerifiedCoding57.6%#82
best: Claude Fable 5 · 95.0%
Terminal-BenchCoding30.0%#12
best: Claude Opus 4.5 (High) · 59.3%
Run it locally
VRAM @ Q4
55 GB
VRAM @ FP16
212 GB
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
~53-55GB minimum at Q4 exceeds a single RTX 4090's 24GB VRAM; needs multi-GPU or CPU offload for full 128K context.
Quantizations
GGUF · AWQ · FP8
GLM-4.5 family
Elo progression across releases
GLM-4.5 (355B-A32B)Jul 20251759.0GLM-4.5-Air (106B-A12B)Jul 20251617.7
Resources
HFzai-org/GLM-4.5-AirGHzai-org/GLM-4.5DOCSdocs.z.ai/guides/overview/pricingAPIAvailable via Zhipu AI / Z.ai (Tsinghua) API
API price $0.2/$1.1 · each benchmark row carries its own source badge (see methodology)