GLM-5: China's Frontier AI Model Built Without Nvidia

GLM-5, Zhipu AI’s 744-billion-parameter open-source model, is the clearest evidence yet that the US chip embargo is not keeping China a generation behind in frontier AI development.

Released on February 11 and built entirely on 100,000 Huawei Ascend 910B processors with zero NVIDIA or AMD silicon, GLM-5 is a 744-billion-parameter model that now outscores GPT-5.2 on two of the most demanding AI benchmarks available. It runs under an MIT license, costs 5 to 6 times less than comparable Western frontier models, and carries a reported hallucination rate lower than any closed-source competitor. The benchmark data has drawn scrutiny, but independent evaluators are confirming the core claims this week.

Benchmark Performance That Changes the Narrative

The numbers are not ambiguous. On SWE-bench Verified, the industry’s hardest real-world software engineering evaluation, GLM-5 scores 77.8 percent compared to GPT-5.2’s 76.2 percent. On Humanity’s Last Exam with tool access, a dataset specifically built to remain unsolvable by near-term AI, GLM-5 posts 50.4 percent against GPT-5.2’s 47.8 percent and Claude Opus 4.5’s 46.2 percent. On BrowseComp, a web research evaluation, GLM-5 scores 75.9 against GPT-5.2’s 72.1.

Claude Opus 4.5 leads on SWE-bench at 80.9 percent, and GPT-5.2 holds the AIME 2025 math benchmark at a perfect 100 percent versus GLM-5’s 88.7 percent. The picture is mixed, as it is with every frontier model. But the ceiling GLM-5 reaches places it squarely in the same tier as models from OpenAI and Anthropic on the tasks that matter most to enterprise buyers: coding assistance, long-horizon reasoning, and autonomous research.

The hallucination data is the most striking claim. Zhipu reports GLM-5 at a 34 percent hallucination rate using its Slime reinforcement learning framework, down from 90 percent on its predecessor GLM-4.7. For context, Claude Sonnet 4.5 is benchmarked at approximately 42 percent on the same evaluation and GPT-5.2 at approximately 48 percent. If those numbers survive independent validation, GLM-5 would have the most reliable factual accuracy of any publicly available model.

Architecture Built for Cost Efficiency

GLM-5 is a Mixture-of-Experts model with 256 total experts and only 8 active per inference pass. The architecture activates just 44 billion of its 744 billion parameters during any given forward pass, keeping compute costs competitive with smaller dense models. The context window extends to 200,000 tokens with a maximum output of 131,000 tokens.

Two borrowed architectural choices stand out. Multi-head Latent Attention cuts memory requirements by 33 percent compared to standard multi-head attention. DeepSeek Sparse Attention, drawn from open-source frameworks pioneered by Chinese AI labs, handles efficient long-context processing without dense attention overhead. Zhipu’s research team pre-trained on 28.5 trillion tokens, a 24 percent increase over the 23 trillion used for GLM-4.5.

The training cluster was 100,000 Huawei Ascend 910B processors running CANN 8.0 and MindSpore 2.5. Zhipu engineers documented scaling challenges including memory bandwidth constraints and inter-node communication latency, but reported that a custom communication library for Ascend-to-Ascend data transfer closed most of the throughput gap with NVIDIA-based systems.

What the Chip Embargo Did and Did Not Accomplish

The US Bureau of Industry and Security has progressively restricted NVIDIA H100, H200, and A100 exports to China since 2022. The restriction forced Chinese AI labs to commit to Huawei Ascend hardware as their primary compute path. The intent was to impose a compute ceiling that would slow frontier model development.

The results are more complicated. NVIDIA-based training retains measurable throughput advantages. GLM-5 took longer and required more engineering overhead to train than equivalent Western models. NVIDIA’s performance lead at scale is real.

But the gap Huawei Ascend needed to close in order for Chinese labs to ship frontier models has been closed. GLM-5 is not a near-miss. It competes on the same leaderboards as the top Western models and in several cases leads them. The embargo accelerated Chinese investment in domestic AI infrastructure rather than preventing frontier development.

The stock market read this clearly. Zhipu shares surged 28.7 percent on the Hong Kong exchange in the days following GLM-5’s release, the company’s largest single-week gain since listing.

Open Source Pricing at Frontier Scale

The commercial implications extend beyond the benchmark sheet. GLM-5 is released under an MIT license, the most permissive open-source license available. Any company can deploy it, fine-tune it, or build products on top of it without licensing fees. Zhipu’s API pricing for GLM-5 runs 5 to 6 times cheaper than comparable GPT-5.2 or Claude Opus 4.5 access.

The combination of frontier benchmark performance, MIT licensing, and sub-parity pricing creates a competitive pressure on Western closed-source labs that model capability comparisons alone do not capture. This is the same dynamic DeepSeek established with its R1 release earlier this year: open-source at frontier quality resets buyer expectations across the entire market.

Enterprise AI procurement teams that were debating between OpenAI and Anthropic are now evaluating a third option that outperforms both on specific high-value tasks, costs a fraction of either, and carries no per-seat licensing restrictions. The benchmark data will be pressure-tested over the coming weeks, but the direction of travel is set.