Zhipu GLM-4.7 Claims Lowest Hallucination Rate and Full Training on Huawei Ascend Without Nvidia

Zhipu AI released GLM-4.7 in May 2026 as an open-weight model carrying two claims that have drawn immediate attention from the research community: a self-reported hallucination rate of 1.2%, the lowest figure published by any frontier lab to date, and a training process conducted entirely on Huawei Ascend 910B chips with no Nvidia hardware involved.

The Hallucination Claim

A 1.2% hallucination rate, if reproducible under rigorous independent evaluation, would represent a substantial advance on the performance of current frontier models, which typically report factual error rates ranging from 3% to 10% depending on the benchmark and measurement methodology. Zhipu has not yet published the evaluation protocol underlying the figure, which means independent verification is not yet possible.

The company's track record on hallucination benchmarks has been respectable — its GLM lineage has consistently performed competitively on TruthfulQA and related evaluations — but the gap between a self-reported 1.2% and externally audited performance is significant. The research community's usual standard is third-party replication using a specified and publicly available benchmark, which for this release has not yet occurred.

The claim is nonetheless commercially relevant. Hallucination is the primary barrier to enterprise deployment of large language models in domains where errors carry material consequences — legal document review, medical information summarisation, financial analysis, compliance checking. A model that demonstrably reduces hallucination rates would capture significant enterprise attention even if the true rate were somewhat higher than advertised.

Training Without Nvidia

The geopolitically notable dimension of GLM-4.7 is the hardware story. The model was developed entirely on a cluster of 100,000 Huawei Ascend 910B chips — no A100s, H100s, H200s, or any other Nvidia hardware in the training stack. This places GLM-4.7 alongside DeepSeek V4 Pro's Ascend-optimised release as evidence that Chinese AI laboratories are making concrete progress toward training pipelines independent of US-origin semiconductor supply.

US export controls on advanced AI chips have been progressively tightened since 2022, with successive restrictions targeting Nvidia's A100, H100, and H800 series. The controls were intended to slow Chinese frontier-model development by removing access to the dominant hardware platform. GLM-4.7's training provenance suggests the strategy is meeting determined technical resistance: Huawei's Ascend architecture, while still trailing Nvidia's best current-generation chips in raw throughput, appears capable of supporting training runs that produce competitive frontier models.

The implications extend beyond China. Any country facing export controls on Nvidia hardware — or choosing to avoid US chip dependencies for sovereign reasons — will note a training infrastructure that produces a credible frontier model as a meaningful proof of concept.

Pricing and Accessibility

At $0.11 per million input tokens, GLM-4.7 is priced at the extreme low end of the frontier-model cost spectrum. Claude Opus 4.7 costs $5 per million input tokens; the ratio is approximately 45 to 1 on input pricing. On output tokens, the differential is comparable.

Zhipu's pricing strategy mirrors DeepSeek's: use aggressive pricing on hosted API access to maximise developer adoption and data collection, while the open-weight release simultaneously builds a community of self-hosting operators who do not pay per token at all. The two distribution models complement each other — hosted access reduces the barrier for experimentation, open weights reduce the barrier for production deployment where data sovereignty or cost at scale matter.

The model is available on Hugging Face under an open licence, though the specific licence terms governing commercial use warrant review by operators intending production deployment. Zhipu has not uniformly used MIT terms across its GLM releases, and licence variations can affect integration into commercial products.

Position in the Open-Source Landscape

GLM-4.7 enters a landscape already reshaped in the past month by DeepSeek V4 Pro and Mistral Medium 3.5. The three releases collectively demonstrate that the gap between proprietary frontier models and openly available alternatives is narrowing on coding benchmarks. Whether the same convergence applies to factual accuracy, instruction following, and multimodal capability — the areas where proprietary models have maintained the largest advantages — will become clearer as independent evaluators process the new releases.

The next significant open-weight milestone to watch is Meta's Behemoth, still in training, which is expected to be the most capable open-weight model released when it ships. Until then, the May 2026 cohort of open releases represents the strongest open-source model generation to date.