Open-source AI model GLM-5.1 with 754B parameters scores 58.4 on SWE-Bench Pro, surpassing GPT-5.4 through 8-hour autonomous coding on Huawei chips.

Open-Source AI Beats GPT-5.4: GLM-5.1’s 8-Hour Coding Feat

by
💡 Key Takeaways

  • China’s Z.ai released GLM-5.1, scoring 58.4 on SWE-Bench Pro to surpass GPT-5.4 (57.7) and Claude Opus 4.6 (57.3) as the world’s top open-source AI coding model.
  • The 754B MoE model was trained entirely on Huawei Ascend 910B chips — zero Nvidia GPUs involved.
  • Z.ai demonstrated 8-hour autonomous coding, building a complete Linux desktop environment through 655 self-directed iterations.

58.4. That single number just rewrote the leaderboard. The score that now sits atop SWE-Bench Pro — the benchmark that measures whether AI can actually fix real-world software bugs — doesn’t belong to OpenAI, Anthropic, or Google.

It belongs to GLM-5.1, a 754-billion-parameter open-source AI model built by Z.ai, a Tsinghua University spinoff that went public in Hong Kong three months ago. The model was released under the MIT license on April 7, 2026, and it was trained entirely on Huawei chips — not a single Nvidia GPU touched the process.

But the benchmark score is only half the story. Z.ai also demonstrated the model coding autonomously for eight straight hours, building a complete Linux desktop environment from scratch without any human intervention. Following DeepSeek V4’s trillion-parameter achievement on Huawei hardware, GLM-5.1 marks the second major proof point that China’s independent AI infrastructure can compete at the absolute frontier.

Open-source AI neural network 3D render (AI-generated)
GLM-5.1’s 754B parameter architecture claims the top spot on SWE-Bench Pro for open-source AI models — Generated by OpenAI DALL-E 3

SWE-Bench Pro 58.4: Dethroning GPT-5.4 and Claude Simultaneously

The numbers are unambiguous. GLM-5.1 scored 58.4 on SWE-Bench Pro, edging past GPT-5.4 at 57.7, Claude Opus 4.6 at 57.3, and Gemini 3.1 Pro at 55.1. It now ranks first among all open-source models globally and third overall — including proprietary systems.

SWE-Bench Pro isn’t a trivia quiz. It pulls real bugs from actual GitHub repositories and requires the model to understand the codebase, identify the root cause, write a fix, and pass the project’s test suite. This is end-to-end software engineering, not code completion.

The model’s capabilities extend beyond pure coding. On the CyberGym benchmark, GLM-5.1 scored 68.7, completing 1,507 real-world cybersecurity tasks from start to finish. That said, a broader composite assessment of total coding capability puts GLM-5.1 at approximately 94.6% of Claude Opus 4.6’s performance — dominant on specific benchmarks, but not uniformly superior across every dimension.

Key point: GLM-5.1 beat GPT-5.4 by 0.7 points and Claude Opus 4.6 by 1.1 points on SWE-Bench Pro — the first time an open-source model has topped proprietary rivals on this benchmark.

754B MoE Architecture: Only 40B Active Parameters Per Token

GLM-5.1’s total parameter count is 754 billion. But the model uses a Mixture-of-Experts (MoE) architecture, meaning only a fraction of those parameters activate for any given input. Specifically, each token processed engages just 40 billion active parameters — roughly 5.3% of the total model.

The practical implications are significant. The model supports a 200,000-token context window and can generate up to 128,000 tokens in a single output. For software engineering tasks, this means GLM-5.1 can ingest tens of thousands of lines of code at once and produce comprehensive modifications in a single pass.

Under the hood, GLM-5.1 employs DeepSeek Sparse Attention (DSA), a technique developed by fellow Chinese AI lab DeepSeek that reduces computational overhead when processing long sequences. The combination of massive parameter count, efficient activation, and sparse attention creates a model that punches well above its inference cost.

In simpler terms: Think of the MoE architecture as a large consulting firm with experts in every field. Instead of putting every consultant on every project, only the relevant specialists are activated for each task — delivering the full firm’s expertise at a fraction of the staffing cost.
Open-source AI chip architecture illustration
The MoE architecture activates only 40B of 754B parameters per token, balancing raw capability with computational efficiency — Photo by Igor Omilaev on Unsplash

8-Hour Autonomous Coding: Building a Linux Desktop Solo

The benchmark scores tell you what GLM-5.1 can do in a controlled test. The 8-hour autonomous coding demonstration tells you what it can do in the wild. Z.ai tasked the model with building a complete Linux desktop environment from scratch, and it executed 655 autonomous iterations — planning, coding, testing, debugging, and optimizing — without a single human intervention.

The velocity of improvement is staggering. Four months ago, the GLM series could sustain roughly 20 autonomous steps before losing coherence or requiring human correction. GLM-5.1 pushed that to 1,700 steps — an 85-fold increase in autonomous endurance.

Z.ai’s leadership sees this as a paradigm shift. The company argues that autonomous work time — not benchmark scores — is now the most important curve in AI development, surpassing even traditional scaling laws.

The demonstration wasn’t limited to building applications. GLM-5.1 also optimized a vector database’s query throughput to 6.9 times the initial production baseline during the same autonomous session. The model executed a full plan-execute-test-fix-optimize loop, proving it can handle not just greenfield development but performance engineering on existing systems.

Key point: Autonomous execution steps jumped from 20 to 1,700 in four months — an 85x increase. Z.ai positions “autonomous work time” as the next critical metric in AI development, beyond traditional scaling laws.

Huawei Ascend 910B: World’s Top Score Without a Single Nvidia Chip

GLM-5.1’s technical achievements carry geopolitical weight. The entire model was trained on Huawei Ascend 910B chips — zero Nvidia GPUs were used at any stage of the process. This makes GLM-5.1 the highest-performing AI model ever built entirely on non-Nvidia hardware.

U.S. semiconductor export controls were designed to constrain China’s AI ambitions by restricting access to cutting-edge GPU technology. GLM-5.1 demonstrates that those constraints have not prevented Chinese labs from reaching — and in specific benchmarks, surpassing — the performance of models trained on the world’s most advanced chips.

Context matters here. Huawei’s Ascend 910B chips still trail Nvidia’s H100 in absolute compute performance. Training GLM-5.1 likely required more chips, more power, and more time than an equivalent Nvidia-based training run would have demanded.

But the end result — the model’s actual capability measured by what it can do — now matches or exceeds Western proprietary models on the benchmarks that matter most for software engineering. The gap between chip performance and model performance has effectively closed.

Server room infrastructure network cables
Trained exclusively on Huawei Ascend 910B chips, GLM-5.1 proves China’s domestic AI infrastructure can compete at the global frontier — Photo by Taylor Vick on Unsplash

A $52.8 Billion Tsinghua Spinoff: The World’s First Public AI Foundation Model Company

Z.ai, formerly known as Zhipu AI, is a Tsinghua University spinoff that listed on the Hong Kong Stock Exchange on January 8, 2026. The IPO raised HKD 4.35 billion (approximately $558 million), making it the first publicly traded AI foundation model company in the world. Its market capitalization currently stands at approximately $52.8 billion.

The valuation belies the balance sheet. Z.ai reported a net loss of CNY 4.7 billion (approximately $683 million) for 2025 — meaning the company burns nearly $700 million annually while commanding a valuation north of $50 billion. The market is pricing in a future that the current financials have not yet delivered.

Investors responded immediately to the GLM-5.1 open-source release. Z.ai’s stock surged 19% intraday on April 8, closing up 15.92%. This pattern of sharp technical-achievement-driven rallies has become characteristic of the stock — a previous earnings release triggered a 35% single-day surge.

MIT License: A 754B Model Anyone Can Use

GLM-5.1 was released under the MIT License on Hugging Face, one of the most permissive open-source licenses available. There are no restrictions on commercial use, modification, or redistribution.

Z.ai also lowered the practical barrier to entry. An FP8 quantized version of the model reduces memory requirements by approximately 50%, making deployment feasible on a wider range of hardware. The model is compatible with major AI development environments including Claude Code, OpenCode, and Kilo Code.

The decision to go fully open-source is strategic, not philanthropic. By building a developer ecosystem around the open weights, Z.ai increases the conversion pipeline to its paid API service while simultaneously benefiting from community contributions that improve the model. In an era where AI energy efficiency has become a critical concern, making a 754B model locally deployable is a powerful ecosystem play.

💡 Related Posts
$5.6M, 1 Trillion Parameters: DeepSeek V4 Drops Nvidia

Price Hikes and the Profitability Paradox: The Open-Source Revenue Equation

Releasing model weights for free while simultaneously raising API prices appears contradictory — but that’s exactly what Z.ai did. Alongside the GLM-5.1 launch, the company increased API prices by 8–17%, the second price hike in 2026.

Current pricing stands at $1.40 per million input tokens and $4.40 per million output tokens. CEO Zhang Peng stated he expects the company’s cloud-based model business to achieve “exponential growth,” citing “continuously strong growth in market demand and rapid increases in user scale” as justification.

Z.ai isn’t alone in this move. Alibaba and Tencent have also raised prices for their AI services, signaling an industry-wide shift across China’s AI sector from growth-at-all-costs to revenue optimization. The question is whether this pricing power can close the gap on $683 million in annual losses quickly enough to justify a $52.8 billion valuation.

💡 Related Posts
$300B AI Investment Surge: 150% Record as 95% See Zero ROI
Key point: Z.ai is executing a dual strategy — open-source weights to capture developers, paid API to monetize usage — while hiking prices for the second time in 2026 alongside industry peers Alibaba and Tencent.

GLM-5.1 is not a single model release — it is a three-part answer. Can open-source AI beat proprietary models at software engineering? Can Huawei chips produce world-class results?

Can an AI system work autonomously for a full workday? On all three counts, the answer is now yes.

The 0.7-point margin over GPT-5.4 on SWE-Bench Pro won’t upend Silicon Valley overnight. But the combination of open-source weights, domestic Chinese hardware, and eight-hour autonomous execution introduces a variable that the AI industry’s current pricing, licensing, and infrastructure assumptions haven’t fully accounted for. Z.ai’s next quarterly earnings and enterprise adoption metrics for GLM-5.1 will determine whether this variable reshapes the competitive landscape or remains a benchmark curiosity.

References

Stay Ahead of the Curve

Get the latest insights on AI, tech, and markets — delivered in both Korean and English.

Explore More Posts