The GLM-5.1 Shock | The Full Picture of the Free MIT-Licensed AI That Beats Opus
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・25分で読める
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"Open-source AI still can't catch up to Claude or ChatGPT" — that conventional wisdom was overturned on April 7, 2026.
GLM-5.1, released by Chinese startup Z.ai, outperformed Opus 4.6 and GPT-5.4 on the most authoritative programming benchmark, and it's available for free to anyone under the MIT License — a truly remarkable move.
This article breaks down what's inside, in plain language.
GLM-5.1 (General Language Model 5.1) is a large language model released on April 7, 2026 by Z.ai (formerly Zhipu AI, an AI startup spun out of Tsinghua University). It is an upgraded version of the previous model GLM-5, with particular enhancements to its capabilities as an "agent" — that is, its ability to think independently and carry out tasks on its own.
The biggest highlight is that it is open source.
The model's internals (weight data) are publicly available on HuggingFace and ModelScope, and anyone can freely download, modify, and use it commercially under the MIT License.
The MIT License is one of the most permissive licenses in the world — essentially "do whatever you want, just keep the copyright notice."
It is the polar opposite of "AI with secret internals" like ChatGPT or Claude.
To use an analogy, it's like "Toyota publicly releasing all of its automobile blueprints for free so anyone can copy them and build an EV." Chinese AI used to carry an image of being "decent but not top-tier," but GLM-5.1 has vaulted straight into the ranks of the world's best.
Z.ai, the developer, was spun out of Tsinghua University's National Key Laboratory of Intelligence Technology and Systems in 2019.
Often called the "Chinese OpenAI," the company rebranded from Zhipu AI to Z.ai in 2025.
It has strong backing from the Chinese government and is one of China's major AI players alongside Huawei and Alibaba.
The main reason GLM-5.1 generated so much buzz is its benchmark results. On SWE-Bench Pro — the most rigorous test for measuring AI programming ability — GLM-5.1 posted an astonishing score of 58.4.
SWE-Bench Pro is an extremely practical test that has AI solve real bug-fix tasks from GitHub.
To pass, an AI must read through hundreds of thousands of lines of code, identify the source of a bug, and write the correct fix.
To put it in terms anyone can understand, it's like "sending an AI into a massive library to find a specific typo and correct it" — that kind of difficulty.
Even more striking is that GLM-5.1 (in reasoning mode) earned 51 points on the "Artificial Analysis Intelligence Index," a comprehensive intelligence metric.
That's far above the average of 27 points for open models of the same size.
It has been hailed as a historic moment — the first open-weight model to stand squarely alongside the leading closed models.
Incidentally, Z.ai officially claims that GLM-5.1 "delivers 94.6% of Claude Opus 4.6's coding performance" at less than one-tenth the price. The developer's dream of "nearly equal performance at one-tenth the cost" is becoming a reality.
GLM-5.1 is a massive model with 744B (744 billion) parameters, but it uses a mechanism called MoE (Mixture of Experts). Think of it as "assembling a team of 256 specialists, then selecting only 8 of them to answer each question."
Another analogy: "A hospital's general reception desk listens to a patient's symptoms and calls in only the 3–4 most relevant specialists for the consultation."
Rather than summoning every doctor, only the necessary ones are activated, which dramatically reduces computational cost.
In practice, only 40B (40 billion) parameters are active during inference.
That's considerably smaller than Claude Opus, which is the secret behind its lower operating cost.
The most groundbreaking feature of GLM-5.1 is its "8-hour autonomous coding" capability. After a human gives an instruction, the AI can continuously cycle through "plan → execute → test → fix → optimize" loops for up to 8 hours with almost no human intervention.
Picture "an excellent intern you brief in the morning and come back to find the work done by evening." Where conventional conversational AI (like ChatGPT) operated as question-and-answer, GLM-5.1 has evolved into the kind of AI you can "hand something off to and trust it to finish."
Another point worth noting is that GLM-5.1 was trained entirely on Huawei Ascend 910B chips. The fact that a cutting-edge AI was built without a single Nvidia GPU has sent shockwaves through the global AI industry.
In other words, it carries significant geopolitical meaning as "China's answer to American semiconductor export controls." For Japan's semiconductor industry as well, this is an important signal that "the era of Nvidia's unchallenged dominance won't last forever."
Let's compare GLM-5.1's positioning against its major rivals.
One key observation: among the top models, GLM-5.1 is the only one using the MIT License. While other open-source AIs often come with restrictions like "commercial use subject to conditions" or "separate fees for large enterprises," GLM-5.1 offers the ultimate freedom: "use it however you like."
The comparison of API pricing makes the impact even more striking.
Opus 4.6 charges $15 per 1M input tokens; GLM-5.1 charges $1.00 — one-fifteenth the price.
With the GLM Coding Plan (promo), you can get started for as little as $3 a month.
That's the price of a single Starbucks latte in exchange for access to the world's most advanced AI coding assistance.
GLM-5.1 was primarily trained on English and Chinese, but it reportedly handles technical instructions in Japanese reasonably well.
That said, there are situations where domestically developed LLMs (such as LLM-jp-4) have the edge in understanding Japanese nuance.
A practical approach would be: "use GLM-5.1 for English-based code generation, and a domestic model for generating Japanese documentation."
For Japan's small and mid-sized IT companies, the arrival of GLM-5.1 represents a major opportunity.
Imagine a web systems development company in Tokyo with 30 employees.
Previously paying ¥600,000 a month for Claude API, they switch to GLM-5.1 and see that cost plummet to ¥60,000.
The ¥540,000 freed up is enough to hire one additional junior engineer.
This kind of shift could begin rippling through Japan's IT industry in the second half of 2026.
That said, caution is warranted.
GLM-5.1 is a product of a Chinese company, and using it via API means data may be transmitted to servers in China.
For sensitive work in government, finance, healthcare, and similar sectors, deploying on your own servers (on-premises) is the standard approach.
By downloading from HuggingFace and running it internally, this risk can be avoided entirely.
There are three main ways to use GLM-5.1.
By subscribing on the Z.ai official website for $3/month (promotional price), you can use GLM-5.1 via a CLI tool, much like Claude.
An integrated coding environment called "GLM Code" is also available, compatible with VSCode and Claude Code.
This is the most accessible option.
Obtain an API key through Z.ai's developer platform "BigModel" and call the model from Python or other languages.
Pricing: input $1.00 / output $3.20 per 1M tokens.
Also available through OpenRouter ($1.26 / $3.96).
For companies handling sensitive data, the best option is to download the weights from HuggingFace (zai-org/GLM-5.1) and run the model on your own servers.
The FP8 quantized version (zai-org/GLM-5.1-FP8) requires significantly less GPU memory and runs practically on around 4–8 NVIDIA H100 GPUs.
Using software called vLLM makes it easy to configure parallel processing across multiple GPUs.
A. In the coding domain, yes.
It claimed first place on SWE-Bench Pro, a highly practical test.
However, Claude and GPT still have an edge in overall capability and conversational naturalness.
For use cases specifically focused on agentic coding and bug fixing, GLM-5.1 can be considered the optimal choice.
A. It functions at a practical level, but there are scenarios where domestically developed models (LLM-jp-4, Sakana Namazu, etc.) outperform it in pure Japanese text generation. For English-based code generation mixed with Japanese comments, it works without issue and is at a level usable by many Japanese development teams.
A. It depends on your use case.
Using it via API carries the possibility that data passes through Chinese servers, so it is not recommended for sensitive data.
On the other hand, since it is published as open source under the MIT License, downloading it to your own servers and running it in a completely offline environment eliminates the risk of data leakage.
The weight data itself can be inspected by anyone, and concerns such as backdoors are being scrutinized by the community.
A. Running the full model (744B) on a personal PC is practically impossible (it would require hundreds of gigabytes of GPU memory).
However, you can access it via API from $3/month, or try it for free through the z.ai chat interface in your browser.
For those who seriously want to run it locally, renting cloud GPU services (such as RunPod or Vast.ai) by the hour is the most realistic approach.
A. Z.ai has indicated a policy of "releasing next-generation models every few months," with hints at the possibility of GLM-6 and a Japanese-specialized model.
Additionally, China's open-source AI contingent — including DeepSeek, Qwen, and Moonshot — are competing fiercely against each other, and even more powerful open models are expected to emerge in the second half of 2026.
The pace of performance improvement in open-source AI is beginning to outstrip that of closed models.
AI is shifting from "an era dominated by a handful of giant corporations" to "an era of open sharing" — and that turning point is happening right now.
Whether you're a developer struggling with AI costs, or a company that has given up on AI because you can't send sensitive data outside the organization, GLM-5.1 offers a new set of options.
Start thinking today about how to incorporate this "free AI that beats Opus" into your projects.
This article is a cross-post from AI Friends.
