← Blog

GLM-5.2 Is Now the Top Scorer Among Open-Weight Models

Mads Kristiansen

CTO, Liviate

China's Z.ai has taken a surprising lead in the open-weights field with GLM-5.2. According to Artificial Analysis, the model is now the leading open-weight model on their Intelligence Index v4.1, scoring 51 — ahead of rivals like MiniMax-M3, DeepSeek V4 Pro max, and Kimi K2.6.

Bigger, longer, and more open

GLM-5.2 launched on June 16, 2026, just 70 days after its predecessor GLM-5.1. The architecture is a Mixture-of-Experts (MoE) model with 744 billion parameters in total, of which only 40 billion are active at a time. The context window has been expanded from 200,000 tokens to 1 million tokens, and the weights are released under an MIT license.

Impressive benchmarks

On Artificial Analysis's primary measure of real-world agentic performance, GDPval-AA v2, GLM-5.2 scores 1,524 — close to proprietary frontier models like GPT-5.5 xhigh (1,514). Other notable results:

Benchmark Score
TerminalBench v2.1 78–81%
GPQA Diamond 89%
HLE 40%
CritPt 21%

The model has made particularly strong gains in scientific reasoning and code. The CritPt score has risen 16 points, and the HLE score is up 12 points.

A Pareto winner — but not the cheapest

GLM-5.2 uses more output tokens per task than its competitors: roughly 43,000 tokens compared to GLM-5.1's 26,000 and DeepSeek V4 Pro max's 37,000. That makes it a relatively expensive open-weight model to run per task — around $0.46 versus DeepSeek's roughly $0.05.

Even so, Artificial Analysis places it on the Pareto frontier: you get more intelligence, but you pay more for it. API pricing is $1.4 / $4.4 / $0.26 per million input/output/cache-hit tokens.

Availability and hosting

The weights are available on Hugging Face, and the model is hosted by a wide range of inference providers: Z.ai, DeepInfra, Novita, Nebius, Parasail, Siliconflow, GMI Cloud, Baseten, and Fireworks.

Open weights as a defensive weapon

In a notable episode, Hugging Face used the open-weight GLM-5.2 to defend itself against an attacker who, according to the company, was using a closed commercial frontier model with refusal restrictions. It's a small but symbolic example of how open weights can give operators more control than closed alternatives.

Why it matters

GLM-5.2 underscores a broader trend: China currently dominates the release of large open-weight models. While American companies like Nvidia, Microsoft, and Meta are now lobbying to protect open weights from overly strict regulation, it's Chinese labs that are actually setting the standard. The question is whether the American industry can keep up — or whether it will instead lock down further models.

Sources

Want to talk about what's actually blocking your AI initiatives?

Mads Kristiansen is happy to have a no-obligation chat.

Book a meeting →