GLM-5 Review: Z.ai's Open-Source 744B Model with Record-Low Hallucinations
GLM-5 from Zhipu AI / Z.ai is a 744B open-source model scoring -1 on hallucination index, 86% GPQA, 92.7% AIME 2026, at $1.00/$3.20 per million tokens.
Zhipu AI, now operating as Z.ai, released GLM-5 on February 11, 2026. It currently ranks #1 among open-weights models on the Artificial Analysis Intelligence Index and holds the top position in both the LMArena Text Arena and Code Arena open-model categories (VentureBeat). Its most distinctive technical result is a score of -1 on the Artificial Analysis Omniscience hallucination index, the best number any model, open or proprietary, has posted on that benchmark.
This review covers the architecture, benchmark results, hallucination data, pricing, and practical use cases.
What is GLM-5?
GLM-5 is the latest generation of Zhipu AI's General Language Model series. It is fully open source under an MIT license, no attribution requirements, no commercial restrictions, unrestricted self-hosting.
The model was developed under the Z.ai branding that Zhipu AI adopted in early 2026 as part of its international expansion. It was trained entirely on domestically produced Huawei Ascend chips, a deliberate signal about China's push toward AI compute self-reliance (NxCode).
Key specs:
- Parameters: 744B total, 40–44B active (MoE)
- Architecture: 256 experts, top-8 activated per token
- Context window: 200K tokens input, 128K output
- License: MIT (fully open source, no attribution required)
- Release date: February 11, 2026
Architecture
GLM-5 uses a Mixture-of-Experts architecture with 256 experts and top-8 activation, meaning 8 experts activate per token, combining to produce approximately 40–44B effective active parameters per forward pass. This is considerably more active capacity than MiniMax M2.5 (10B active) or Qwen3-235B (22B active), which partly explains GLM-5's stronger performance on knowledge-intensive tasks.
The training recipe uses a technique called SLIME (asynchronous reinforcement learning), developed by the Zhipu team. SLIME decouples the data collection and policy update phases of RL training, allowing more efficient use of GPU clusters during the alignment phase. The technical paper is on arXiv (arxiv.org/html/2602.15763v1).
Pre-training used a 28.5 trillion token dataset, large by current standards and likely responsible for GLM-5's strong factual knowledge performance.
Benchmark performance
Mathematics
AIME 2026 (the American Invitational Mathematics Examination, most recent year) is one of the harder math benchmarks in current use because models cannot overfit to it through training data. GLM-5 scores 92.7%: among the best results recorded by any model at time of release.
Science reasoning
GPQA Diamond, testing graduate-level biology, chemistry, and physics, comes in at 86.0%. This places GLM-5 above MiniMax M2.5 (84.8%) and at a level that was the exclusive territory of the most expensive proprietary models six months ago.
Software engineering
SWE-Bench Verified: 77.8%. Strong but not the leader in this category, MiniMax M2.5 (80.2%) and Kimi K2.5 (76.8%) are its immediate competitors. GLM-5 is a capable all-rounder for coding rather than a specialist.
Speed
Inference speed: 72.5 tokens per second, with a time-to-first-token of 1.30 seconds. Fast enough for interactive use without batching tricks.
Agentic benchmarks
GLM-5 leads open-source models on BrowseComp, Vending Bench 2, and MCP-Atlas according to the Zhipu release notes. These agent-oriented benchmarks test the model's ability to use tools, navigate multi-step tasks, and retrieve information from web searches, increasingly the relevant metrics as developers build autonomous pipelines rather than single-turn assistants.
Artificial Analysis Intelligence Index
GLM-5 ranks #1 among open-weights models on the AA Intelligence Index and holds the top position in both LMArena Text Arena and Code Arena for open models.
Hallucination reduction
The most headline-grabbing result from the GLM-5 release is its score of -1 on the AA-Omniscience Index.
The Omniscience Index measures calibrated hallucination: negative scores indicate the model is less likely to confabulate false information than the average model in the benchmark pool. A score of -1 is the best result recorded across all models tested, GLM-5 produces fewer confident wrong answers than any other model on the benchmark, including top proprietary systems.
Compared to its predecessor GLM-4.7, this is a 56% reduction in hallucination rate (Artificial Analysis, GLM-5). In practical terms: GLM-5 is significantly less likely to invent citations, fabricate statistical figures, or state wrong facts with false confidence.
This matters in real deployments. Hallucinations are one of the primary reasons AI assistants fail in high-stakes workflows, a factual error in a legal document, a fabricated drug interaction in a medical summary, or a wrong API endpoint in generated code can cause serious problems. A model that knows what it does not know is more reliable in these settings, even when its average benchmark score is not the absolute highest.
The SLIME reinforcement learning technique and the large pre-training corpus are both likely contributors to this result. Better calibration generally follows from more RL feedback aligned with honesty objectives and a richer knowledge base that reduces uncertainty.
Pricing
GLM-5 is available through the Z.ai API at:
| Tier | Price per 1M tokens |
|---|---|
| Input (standard) | $1.00 |
| Input (cached) | $0.20 |
| Output | $3.20 |
For context: Claude Opus 4.6 costs $5.00 input / $25.00 output. GLM-5 is approximately 5x cheaper on input and 8x cheaper on output while ranking above Claude Opus on some benchmarks (Z.ai pricing docs).
The 200K context window combined with the $0.20/M cached input price makes it economical to run large documents repeatedly, a cached 100K-token document costs $0.02 to re-reference, which is practically free at scale.
Note: Zhipu raised GLM-5 pricing by approximately 30% shortly after launch due to high demand, suggesting the initial pricing was partly promotional. The figures above reflect post-adjustment pricing (TechLoy).
Use cases
Factual knowledge pipelines where accuracy matters
The -1 Omniscience score makes GLM-5 the strongest choice for applications where hallucinations are particularly costly: legal research, medical information synthesis, scientific literature review, financial analysis. It is better calibrated than anything else in the open-weights tier and competitive with proprietary systems on this metric.
Mathematics and quantitative reasoning
A 92.7% AIME 2026 score indicates genuine mathematical reasoning ability. GLM-5 is well-suited for tutoring platforms, quantitative finance tools, and engineering calculation support.
Long-document analysis
The 200K input / 128K output window, combined with the cached input pricing of $0.20/M, makes GLM-5 practical for workflows that repeatedly reference long documents. Contract analysis, regulatory compliance checking, and systematic literature review all benefit from this combination.
Agentic and tool-use applications
GLM-5's top placement on BrowseComp, Vending Bench 2, and MCP-Atlas indicates strong performance in multi-step tool use. For developers building research agents, automated data collection pipelines, or orchestrated multi-step workflows, GLM-5 is a strong open-weights candidate.
Self-hosted deployments with no licensing restrictions
The MIT license with no attribution requirements means teams can deploy GLM-5 weights on their own infrastructure, integrate it into commercial products, and modify the model without any legal friction. This is the most permissive license among the major Chinese open-weights frontier models.
Limitations
Output length cap
The 128K output token limit is lower than some competitors (Kimi K2.5 supports 256K total context). For workflows that require extremely long generation, complete software projects in one pass, book-length writing, this may be a constraint.
Coding is strong but not the leader
At 77.8% on SWE-Bench Verified, GLM-5 is a capable coding model. But MiniMax M2.5 scores 80.2% on the same benchmark. Teams whose primary use case is autonomous software engineering should run their own evals against both models before committing.
Writing quality data is limited
GLM-5 does not appear prominently in lechmazur's creative writing rankings as of this writing. The model's strength is clearly in factual, technical, and mathematical domains. For prose quality and creative tasks, other models in the same price range may produce better results.
Huawei Ascend training infrastructure
Training on Huawei Ascend chips rather than NVIDIA is a technical novelty, but it also means the training infrastructure is less battle-tested at the frontier. Whether this produces subtle quality differences in practice is not yet well-documented in third-party evaluation.
Verdict
GLM-5 occupies a distinctive position in the open-weights world: it is the most reliable model available by hallucination metrics, ranks #1 among open models on the AA Intelligence Index, and ships under a fully unrestricted MIT license.
If your primary concern is factual accuracy, minimizing confident wrong answers in production, GLM-5 is the correct choice. No other open model matches it on the Omniscience benchmark, and it is competitive with or better than the best proprietary models on that dimension.
For coding-first applications, MiniMax M2.5's 80.2% SWE-Bench score and lower output price give it an edge. For agentic web research, Kimi K2.5's BrowseComp swarm results are stronger. But for any team that needs high-confidence factual outputs, mathematical reasoning, or science-domain question answering, and wants the freedom of a true MIT license, GLM-5 is the model to evaluate first.
Available on AfricanAI: test GLM-5 alongside other frontier models without managing API keys.
Sources: