TL;DR
- GLM-5, a 744-billion-parameter mixture-of-experts model with 40 billion active parameters, was released as the most significant open-source model of 2026 — trained entirely on Huawei Ascend chips without any Nvidia hardware.
- The model was trained on 28.5 trillion tokens and uses DeepSeek Sparse Attention architecture, with GLM-5.1 achieving 94.6% of Claude Opus 4.6’s coding performance.
- Pricing sits at $3/month compared to Claude Max’s $100-200/month, marking a 30x cost advantage while delivering near-frontier performance.
- The release reshapes hardware dependency assumptions and raises questions about U.S. export controls and geopolitical AI infrastructure competition.
Huawei Ascend Chips Just Trained a Frontier-Class Model
GLM-5 dropped in April 2026 as the year’s most important open-source model release — and it wasn’t trained on a single Nvidia chip. According to BuildFastWithAI, the 744-billion-parameter mixture-of-experts model was trained on 28.5 trillion tokens using exclusively Huawei Ascend hardware, marking the first time a frontier-class model has bypassed Nvidia’s infrastructure entirely.
The model activates 40 billion parameters during inference despite its 744-billion total parameter count, using a Mixture of Experts architecture combined with DeepSeek Sparse Attention. GLM-5.1, a refined version focused on coding tasks, hits 94.6% of Claude Opus 4.6’s capability — a benchmark that would’ve been unthinkable for open-source models 18 months ago.
BuildFastWithAI called it bluntly: “GLM-5 is the most important open-source release of 2026…built on Mixture of Experts with DeepSeek Sparse Attention, trained on 28.5 trillion tokens — entirely on Huawei Ascend chips without Nvidia hardware.”
Why GLM-5’s Huawei Training Infrastructure Rewrites the Playbook
This isn’t just another model release. It’s a proof point that Nvidia’s stranglehold on AI training infrastructure has cracked.
For years, the assumption was simple: if you wanted to train a frontier model, you needed Nvidia H100s or their successors. Export controls targeting China reinforced that narrative — cut off access to advanced chips, the thinking went, and you cut off access to frontier AI capabilities. GLM-5 torches that assumption.
Training a 744-billion-parameter model on 28.5 trillion tokens requires staggering compute. That Huawei’s Ascend chips handled the entire training run — without Nvidia fallback — signals that alternative chip architectures have matured faster than most Western observers expected. And it raises uncomfortable questions about whether export controls are solving the problem they were designed to address or simply accelerating indigenous chip development.
The pricing tells you everything about the strategic intent here. At $3/month, GLM-5 undercuts Claude Max’s $100-200/month pricing by a factor of 30 while delivering 94.6% of the coding performance. That’s not a marginal improvement. That’s a market-shaping wedge.
I’ve watched open-source models chase proprietary frontiers for a decade, and the gap has never closed this fast. GLM-5 doesn’t just narrow the gap — it collapses it in specific domains while making deployment economically trivial. If you’re a developer choosing between paying $200/month for Claude Max or $3/month for GLM-5.1 with 94.6% of the capability, the math isn’t subtle.
Think of it like this: Nvidia’s chip dominance was a dam holding back a reservoir of global AI talent and capital. Export controls were supposed to reinforce the dam. Instead, they’ve forced engineers to dig channels around it — and one of those channels just proved it can carry the same water volume.
Open-Source Models Are Now Within Striking Distance of GPT-5.4
GLM-5 doesn’t exist in a vacuum. It’s the most visible data point in a broader trend that’s been building since late 2025: open-source models are catching frontier proprietary systems across the board.
Meta’s Llama 4, DeepSeek V4, and a constellation of community models now sit within a few benchmark points of GPT-5.4 and Gemini 3.1 Pro. The gap that once measured in years now measures in months — or in some domains, in single-digit percentage points. As of April 2026, the capability delta between the best open-source models and the best proprietary models has effectively vanished for many production use cases.
That convergence reshapes the competitive landscape. Proprietary model labs can’t rely on a sustained capability moat anymore. They’re competing on deployment speed, API reliability, safety guarantees, and ecosystem lock-in — not raw intelligence. And open-source models like GLM-5 are competing on cost, customizability, and infrastructure independence.
For developers, this is the inflection point. You’re no longer making a capability trade-off when you choose open-source. You’re making a strategic choice about vendor lock-in, cost structure, and geopolitical risk.
What Happens When U.S. Export Controls Meet Huawei-Trained Models?
But here’s where it gets messy. GLM-5’s reliance on Huawei Ascend chips puts it squarely in the crosshairs of U.S. export control policy.
If American companies or researchers want to use GLM-5 in production, they’re deploying a model trained on hardware from a company explicitly targeted by U.S. sanctions. That doesn’t violate export controls directly — the model weights themselves aren’t restricted technology — but it creates a legal and reputational gray zone that enterprises will need to navigate carefully.
Does using a Huawei-trained model constitute indirect support for a sanctioned entity? Will U.S. regulators treat GLM-5 differently than models trained on Nvidia or AMD hardware? These aren’t hypothetical questions. They’re the kind of compliance headaches that slow enterprise adoption, even when the technology is superior.
And there’s a second-order effect: if Huawei-trained models become the default in open-source AI, Western companies face a choice between cutting-edge capabilities and geopolitical alignment. That’s not a technical problem. It’s a strategic one.
Three Things to Watch as GLM-5 Rolls Out
First, monitor whether U.S. or European regulators issue guidance on deploying models trained on Huawei hardware. If they don’t, silence is tacit permission. If they do, the guidance will shape which companies can adopt GLM-5 and under what conditions. Enterprise legal teams are already gaming this out — watch for public statements from major cloud providers about their stance on Huawei-trained models.
Second, track whether other Chinese AI labs follow GLM-5’s lead and publicize Ascend-based training runs. If GLM-5 is an outlier, it’s a curiosity. If it’s the first of many, it’s a paradigm shift. DeepSeek, Alibaba, and Baidu all have the resources to train frontier models on domestic hardware — whether they choose to advertise that fact will signal how the industry is positioning itself relative to export controls.
Third, watch the benchmark wars heat up. GLM-5.1 hit 94.6% of Claude Opus 4.6 on coding tasks, but that’s one domain. If subsequent releases close the gap on reasoning, multimodal understanding, and long-context tasks, the proprietary model labs lose their last defensible moat. The next six months will show whether GLM-5’s performance is a ceiling or a floor.
FAQ
What makes GLM-5 different from other open-source models?
GLM-5 is a 744-billion-parameter mixture-of-experts model trained entirely on Huawei Ascend chips without any Nvidia hardware, using 28.5 trillion tokens. It activates 40 billion parameters during inference and delivers near-frontier performance at $3/month — a 30x cost advantage over proprietary alternatives like Claude Max.
How does GLM-5’s coding performance compare to Claude Opus?
GLM-5.1, the coding-focused version, achieves 94.6% of Claude Opus 4.6’s capability on coding benchmarks. That’s close enough to frontier performance that most developers won’t notice the gap in production use, especially given the massive price difference.
Are there legal risks to using a model trained on Huawei chips?
Using GLM-5 doesn’t directly violate U.S. export controls, but it creates a gray zone for American companies. The model was trained on hardware from a sanctioned entity, which could raise compliance questions for enterprises operating under strict regulatory oversight. Legal guidance hasn’t been issued yet, so companies are navigating uncharted territory.
Does GLM-5 prove export controls on AI chips aren’t working?
GLM-5 demonstrates that China’s domestic chip industry can train frontier-class models without Nvidia hardware, which undermines one goal of export controls. Whether that means the controls have failed depends on your definition of success — they’ve clearly accelerated indigenous chip development rather than blocking AI progress outright.
Source: BuildFastWithAI
