Chinese Open-Source AI Just Beat GPT-5.4 on Coding

On the same day Anthropic announced it was gating its best model to 50 companies, Zhipu AI did the exact opposite. The Beijing-based lab open-sourced a frontier-class model that outperforms GPT-5.4 on coding benchmarks. Weights freely downloadable. Permissive license. No gated access, no waitlist, no enterprise contract. As a practitioner I have to be honest: the best open-weight model in the world right now comes from China, and it's not particularly close.

This is worth paying attention to whether you're building AI in Silicon Valley, Shenzhen, or anywhere else. The open-source AI center of gravity has shifted east, quietly, over the last eighteen months. The four labs leading it — Zhipu, DeepSeek, Alibaba's Qwen team, and Moonshot AI — each ship on a cadence and a quality level that most Western open-weight efforts can't match.

What Zhipu Just Shipped

Zhipu AI released GLM-5 on April 7, 2026 under a commercially permissive license. The headline number is that on coding evaluations — specifically HumanEval, SWE-bench Verified, and MBPP — the model scores higher than GPT-5.4. Not marginally. By a clear margin, on benchmarks that the major labs themselves treat as definitive.

The release includes full weights, a reference inference setup, and an evaluation harness anyone can rerun. Zhipu has made it easy to verify the claims, which is the opposite of how most benchmark disputes go. You can download the model, run it on the same tasks, and see the same numbers.

The Four Labs Shaping Open-Weight AI

GLM-5 is the newest entry, but it isn't isolated. Four Chinese labs have, between them, shipped most of the strongest open-weight releases of the last 18 months. Each has a distinct technical personality and each is worth knowing.

None of these labs are a monolith. They compete with each other. They publish conflicting results. They disagree on architecture. The thing they share is a commitment to open weights at a quality tier Western labs have largely walked away from.

The Great Open/Closed Split

Zoom out and the pattern is striking. In April 2026, the same week Anthropic gated Claude Mythos to 50 companies, Zhipu open-sourced a coding-frontier model and Google released Gemma 4 under Apache 2.0. Open and closed are pulling apart — and geographically, the open side is increasingly Chinese.

Both strategies are rational. Both are working — for different definitions of winning. The closed strategy maximizes revenue per user and concentrates capability. The open strategy maximizes reach and builds an ecosystem the originating lab can keep extending. Which "wins" depends entirely on what you think matters in the long run.

How to Actually Use GLM-5

If you want to try it today, the fastest path is Hugging Face. Zhipu published the weights alongside the release announcement. For a quick test on a single GPU, pull the quantized variant and run it through vLLM:

from vllm import LLM, SamplingParams

llm = LLM(
    model="zhipuai/glm-5-coding",
    dtype="bfloat16",
    tensor_parallel_size=1,
)

prompt = "Write a Python function that takes a directory and " \
         "returns the 10 largest files recursively, excluding " \
         "dotfiles. Include docstring and type hints."

out = llm.generate(
    prompt,
    SamplingParams(temperature=0.2, max_tokens=600),
)
print(out[0].outputs[0].text)

That's the full setup. No API key, no credit card, no quota. It runs on hardware you already have. For coding workflows specifically — IDE integrations, PR review agents, code search — this is genuinely competitive with what you'd get from a paid GPT-5.4 endpoint, at zero per-call cost.

What Builders Should Actually Think About

The Bottom Line

Models don't have nationalities in any way that matters to the code you ship. They have weights, licenses, and benchmarks. GLM-5 has the best of all three for coding workflows right now, and it's free. Go pull it, run it on your stack, and see what it can do. That's the practitioner's job, regardless of which city the weights came out of.