Z.ai launches GLM-4.5 to supercharge coding and AI agents

Most open-source labs right now are locked in a race to build wider MoE models with dozens of experts. Z.ai is taking the opposite route.

The team just released its new flagship series, GLM-4.5 and GLM-4.5-Air, alongside FP8 variants and day-one deployment support on SGLang. Instead of chasing sheer scale, they redesigned the internals around deep reasoning and agentic workflows.

GLM-4.5 lands at 355B total parameters with only 32B active, while the smaller GLM-4.5-Air sits at 106B total with just 12B active.

Going deep instead of wide

While rivals like DeepSeek-V3 and Kimi K2 lean heavily into width, Z.ai decided to prioritize depth over width. That means fewer experts and smaller hidden dimensions, but significantly more transformer layers to tackle hard reasoning problems.

The team packed 96 attention heads into a 5120 hidden size—roughly 2.5 times denser than normal—supported by grouped-query attention, QK-Norm, and the Muon optimizer. For inference speed, they added a Multi-Token Prediction (MTP) head to enable speculative decoding right out of the box.

Across 12 comprehensive benchmarks spanning agents, coding, and reasoning, Z.ai claims GLM-4.5 ranks third overall among frontier models, while the compact Air version sits at sixth.

Its agent benchmarks stand out the most. The flagship scored 90.6% on tool-calling success rate, reached 64.2 on SWE-bench Verified, and scored 26.4% on BrowseComp, edging out Claude 4 Opus. In head-to-head coding evaluations, Z.ai reports a 53.9% win rate over Kimi K2 and 80.8% over Qwen3-Coder.

Why it matters

If you are building autonomous coding agents or enterprise tool-use workflows, GLM-4.5 is worth a serious look today.

The primary barrier to deploying open-source agents has rarely been pure token speed; it has been reliability across multi-turn function calls. Pairing a 90.6% tool-calling rate with ready-to-run SGLang scripts and FP8 checkpoints makes self-hosting frontier-grade agents far less painful.

We think Z.ai made a very smart architectural trade-off here. By keeping active parameters down to 32B and 12B, they made high-performance reasoning runnable without requiring massive hyperscaler infrastructure. It puts direct pressure on proprietary API providers whose margins depend on charging a premium for complex agent reasoning.

That said, deeper networks carry inherent latency trade-offs during sequential inference. Speculative decoding helps, but GLM-4.5 will still need to prove in real-world messy codebases whether its benchmark wins translate to consistent daily developer workflows.