GLM-4.7: The Open-Weight Model Claiming 1.2% Hallucination — Built Without a Single NVIDIA GPU

Two stories in one release

When Zhipu AI published GLM-4.7 in May 2026, it bundled two genuinely distinct stories into a single model card. The first story is about hallucination reduction: the lab claims a 1.2% error rate on factual recall tasks, a figure that, if it holds up under independent scrutiny, would set a new bar for open-weight models. The second story is about hardware: every training step ran on Huawei Ascend 910B and 910C chips, with no NVIDIA silicon anywhere in the pipeline. Both stories carry significant implications for builders in India, the UK, and anywhere else where inference supply chains and data sovereignty are live concerns.

GLM-4.7 is a different product from its stablemate. Zhipu's agentic flagship for software development is GLM-5.1, which we covered in April alongside Llama 4 and Mistral Small 4 in our roundup of open-weight models that finally caught up with proprietary alternatives. GLM-4.7 is narrower in ambition: it is targeted specifically at workloads where factual accuracy is the primary constraint — summarisation, question answering, document retrieval, legal and financial analysis — rather than multi-step agentic execution.

What does "hallucination rate" actually mean?

The phrase "hallucination rate" is widely used and almost never precisely defined in the same way twice. Before building anything production-critical on GLM-4.7's headline figure, you need to understand what Zhipu measured and, crucially, what it did not measure.

A hallucination in language model evaluation typically refers to a factual claim the model generates that is either demonstrably false or unverifiable against the source material. In retrieval-augmented generation (RAG) systems, a hallucination usually means the model invented a fact not present in the retrieved context — sometimes called a "faithfulness failure." In open-domain question answering, it means the model stated something that contradicts established knowledge.

The most common benchmark suite used to measure this is TruthfulQA, which covers 817 questions across 38 categories where humans commonly hold false beliefs. A model that mimics human tendencies towards confident wrong answers scores badly here. A model trained specifically to hedge or refuse on uncertain ground scores better, regardless of whether it has better underlying knowledge. Zhipu AI has not published full methodology details at the time of writing, but the 1.2% figure appears to come from a combination of TruthfulQA-style factual recall tasks and Zhipu's proprietary hallucination benchmark run on Chinese-language corpora.

That said, even a lower bound of 1.2% is meaningful. Competing open-weight models at the time of writing cluster between 3% and 8% on comparable evaluation suites. If GLM-4.7 sustains even half of that improvement on independent evals, it represents a material advance for any workload where a wrong answer is more costly than no answer.

The Huawei Ascend story: why the hardware matters

The training hardware story is arguably more consequential than the benchmark. Since October 2022, the US Bureau of Industry and Security has progressively tightened export controls on advanced semiconductor chips to China. The A100, H100, and their successors are all restricted. The direct consequence was supposed to be a significant slowdown in frontier AI development in China. GLM-4.7 is part of a growing body of evidence that this assumption is incorrect.

Huawei's Ascend 910B and 910C chips are fabricated by SMIC using processes that trail TSMC's leading edge by roughly two to three generations, but they are purpose-built for large-scale transformer training workloads. Zhipu's engineering team has been transparent about running the full GLM-4.7 training run on Ascend clusters, without any NVIDIA hardware in the pipeline. The model achieves frontier-quality results on factual accuracy despite training on silicon that, on raw FLOP-per-second measurements, underperforms H100s significantly. This implies Zhipu has invested heavily in training efficiency — better data curation, improved learning rate schedules, and likely reinforcement learning from human feedback pipelines that are more targeted than those used in earlier GLM generations.

For builders outside China, the immediate practical implication is that GLM-4.7's inference does not require Ascend hardware. The weights are open and run on standard transformer inference stacks — vLLM, llama.cpp, and Hugging Face Transformers all support the GLM architecture. The Ascend training story matters for two other reasons: first, it demonstrates the viability of a non-NVIDIA AI supply chain at frontier quality; second, it raises questions about long-term inference pricing, since a lab that trained on non-NVIDIA silicon has a structurally different cost basis for compute.

Pricing in context: $0.11 per million tokens

At $0.11 per million input tokens, GLM-4.7 is among the cheapest frontier-quality inference available via API today. For context, as of May 2026, GPT-4o-mini runs at approximately $0.15 per million input tokens and Gemini Flash 2.0 at roughly $0.10. DeepSeek V4 has driven aggressive price compression across the sector, and GLM-4.7 slots into that competitive band.

The cost story matters most for high-volume, hallucination-sensitive workloads. If you are running a document summarisation pipeline over tens of millions of tokens per day, the difference between $0.11 and $0.15 per million tokens compounds significantly. At 100 million tokens per day, that is a saving of roughly $1,460 per month before any volume discounts. AI inference costs are falling 95% per year — but until your specific workload is at the bottom of that curve, every basis point of cost reduction is worth evaluating.

For teams running layered caching architectures, GLM-4.7's low cost baseline makes it an attractive candidate for the "uncached miss" slot in a semantic cache pipeline. We covered how to cut LLM API costs 70-90% with layered caching in production — the same architecture applies here, with GLM-4.7 as the upstream model.

Comparison: GLM-4.7 vs the competitive field

The table below summarises the key dimensions for production decision-making. Hallucination rates are lab-reported unless otherwise noted. Context window figures are maximums; practical usable context (before quality degradation) is typically 40–60% of the stated maximum for most models.

The standout observation from this table is that the three cheapest models — GLM-4.7, Gemini Flash 2.0, and DeepSeek V4 — are all strong performers on hallucination benchmarks. The era of assuming that low cost implies higher error rates is over. The tradeoff has shifted: what you give up in the sub-$0.15/MTok band is not necessarily accuracy, but ecosystem maturity, fine-tuning tooling, and — for proprietary models — enterprise support contracts.

Where GLM-4.7 fits in a builder's model portfolio

GLM-4.7 is not a general-purpose workhorse in the way GPT-4o-mini or DeepSeek V4 are positioned. Its design choices point clearly to a specific set of use cases where factual precision outweighs creative flexibility or multi-step reasoning depth.

The workloads where GLM-4.7 is worth evaluating first: document-grounded question answering where the model must cite only what the retrieved context contains; legal and regulatory summarisation where fabricated precedents or case references carry real liability; financial data extraction from earnings reports, annual accounts, or regulatory filings; medical literature synthesis where citation accuracy is a patient safety concern; and customer-facing FAQ systems where a confident wrong answer erodes trust more rapidly than a correct hedge.

The workloads where you should look elsewhere: long-horizon coding agents (use GLM-5.1 or DeepSeek V4); creative writing and brand voice generation where constraint and precision actively impede output quality; multi-modal tasks (GLM-4.7 is text-only at launch); and any workload requiring a context window beyond 128K tokens, where Gemini Flash 2.0's one-million-token window is a qualitatively different capability. Gemma 4's configurable thinking modes are also worth evaluating if your workload benefits from multi-step chain-of-thought reasoning at the open-source price point.

Supply chain implications for India and the UK

The Ascend training story carries different weight depending on where you are building.

For builders in India, the US chip export control regime operates through the BIS Entity List and Validated End User authorisation requirements. Indian hyperscale cloud providers have navigated these regulations to offer NVIDIA-based inference services, but the supply chain remains dependent on US export policy decisions. GLM-4.7's existence demonstrates that a frontier-quality model can be produced and offered for inference without touching NVIDIA hardware at any point in its production. This matters for any builder evaluating long-term vendor risk in their AI infrastructure stack.

For builders in the UK, the relevant frame is data residency and non-hyperscaler inference options. UK financial services regulators (FCA, PRA) and NHS digital standards increasingly require evidence that AI systems used in regulated contexts can demonstrate supply chain transparency. A model whose training provenance is documented and whose inference can be run on-premises — on standard x86 or ARM servers using vLLM — offers a different compliance posture than one locked to a US hyperscaler's managed API.

Neither of these arguments is a reason to deploy GLM-4.7 without evaluation. They are reasons to include it in your evaluation shortlist if supply chain sovereignty is a live concern in your procurement process. The Verified AI Builders network includes practitioners who have navigated exactly these procurement decisions across both markets — worth consulting before committing to an inference architecture.

Builder's evaluation checklist

The competitive trajectory: what comes next

GLM-4.7 arrives at a moment when hallucination reduction has become the primary battleground for the second tier of frontier models — the sub-$0.20/MTok class that is increasingly where real production workloads land. The first wave of open-weight competition, which we tracked in our open-source AI news coverage throughout Q1 2026, was about raw capability: who could match GPT-4 on MMLU, who could crack SWE-bench, who could run locally on a Mac. That race is largely settled. The models that survive in production over the next 12 months will be distinguished by reliability, not capability.

Reliability in LLM systems has three components: factual accuracy (hallucination rate), consistency (low variance across equivalent prompts), and graceful failure (well-calibrated refusals). GLM-4.7 is the first open-weight model to make a public bet that hallucination reduction is the primary product differentiator. If independent evaluations confirm even 60–70% of the claimed improvement over competitors, it will shift the evaluation criteria for builders across the industry.

Zhipu's roadmap implies continued narrowing of the gap between GLM-4.7's factual precision specialisation and GLM-5.1's agentic breadth. The two-model strategy — a precision-focused model and an agentic flagship — mirrors what we see from other frontier labs. The next version of GLM-4 is likely to close the context window gap with Gemini Flash while maintaining the hallucination rate advantage. Watch this space.

For practical guidance on fitting models like GLM-4.7 into a cost-efficient inference architecture, our analysis of how falling inference costs enable profitable AI products lays out the framework for deciding when to host, when to use managed APIs, and how to layer models by cost and capability.