Can you own it?

Intended & out-of-scope use

GLM is Zhipu AI's open-weight family - Zhipu is the Beijing company (spun out of Tsinghua's THUDM lab) that ships internationally under the Z.ai brand. The current flagship GLM-4.6 is a 355B-total / 32B-active Mixture-of-Experts model with a 200K context window, tuned for coding and agentic use, alongside GLM-4.5 (355B/32B), the lighter GLM-4.5-Air (106B/12B), and the dense GLM-4-32B-0414 and GLM-4-9B-0414.

Intended use is general-purpose assistant work, with a clear tilt toward coding, tool use, and agentic workloads where GLM-4.6 is strongest. Deploy the instruct checkpoints behind your own guardrails. Out of scope without additional controls: EU high-stakes use of the 355B MoE while the systemic-risk question is unresolved (see EU AI Act below), and any deployment that assumes MIT terms without checking the per-checkpoint LICENSE - the original glm-4-9b used a custom non-OSI license.

Known limitations, bias & failure modes

• China-aligned alignment and hosted-API moderation. The hosted Z.ai API applies Chinese content moderation on politically sensitive topics, and the open weights carry China-aligned alignment. Account for this behaviourally rather than assuming Western-lab defaults.
• Lighter safety tuning. Instruct alignment withstands casual jailbreaks but is lighter than the large Western labs, with no companion guard/classifier model shipped.
• Capability is domain-concentrated. Leadership is strongest in coding/agentic tasks rather than uniform across all domains.
• Opaque training data. The corpus is not released, so you cannot inspect it for known problem sources - you are reasoning about an opaque artifact.

Openness tier & components

GLM is open_weights, not open-science. Open: downloadable weights, model cards, released inference code, and technical reports. Closed: the training data and the training pipeline. Partial: evaluation is only partially reproducible. The single most important openness fact is the license: the GLM-4.5/4.6 and 0414 series ship under a genuine OSI-approved MIT license (ev-license-mit) - a real advantage over the restrictive community licenses some China peers use - but this is not backed by an independent openness classification here, so Dimension 1 rests on publisher evidence and scores 3.

License terms & permitted use

The GLM-4.5, GLM-4.6, and GLM-4-0414 series are MIT - OSI-approved, with no field-of-use restriction and unconditional commercial use. You may use, modify, redistribute, and commercialize derivatives subject only to the MIT attribution terms.

Verify the LICENSE per checkpoint. Licensing is per-variant: the newer families are MIT, but the original glm-4-9b shipped under a custom, non-OSI "glm-4" license (ev-license-split). Do not assume MIT across every GLM checkpoint - read the LICENSE file on the exact model you pull. This clean-but-conditional picture is a load-bearing input to the legal score (3).

Supply-chain & provenance

Weights are distributed from the verified zai-org org on Hugging Face in safetensors with per-file checksums, and dual-published on Zhipu's ModelScope (ev-modelscope). Two supply-chain surfaces to note: the dual-hub distribution (verify checksums match across Hugging Face and ModelScope) and the legacy THUDM org, where older GLM/ChatGLM checkpoints still live - the current canonical org is zai-org. The checkpoint trust checklist scores 5/8; there is no cryptographic weight signing or SLSA attestation, which is why provenance is a 4, not a 5. No incidents are on record for the canonical org. Pin the exact revision and verify checksums on download.

EU AI Act posture

GLM is a GPAI model. On license grounds the MIT-licensed sizes are genuine free/open-source releases with public parameters and usage information, so they plausibly qualify for the Article 53 open-source exemption from the Annex XI/XII technical-documentation duties. Two caveats keep this at partial:

1. Systemic risk is unresolved. The 355B GLM-4.5/4.6 MoE may approach the 10²⁵-FLOPs threshold, but this is not publicly confirmed. If it crosses, the exemption is void for that model and the full Article 55 package (not published) would be owed. The smaller dense 9B/32B are well under.
2. Surviving obligations are unmet. A copyright policy and a public training-content summary survive the exemption, and neither is published. Because the training corpus is not released, a downstream provider cannot assemble the training-content summary from upstream artifacts - unlike a fully-open family.

A China-based provider is unlikely to furnish an EU AI Office documentation package, so an EU deployer inherits a clean MIT license and a usable model card but must self-assemble the compliance material and resolve (or avoid) the systemic-risk question for the 355B MoE.

Benchmarks & evaluation

GLM-4.6 evaluates as strong on real-world coding, tool use, and agentic tasks and competitive in general capability - its coding/agentic strength is the standout. OneHill did not run its own benchmarks this session; the picture is aggregated from third-party evaluation (ev-perf-coding). Treat published scores as third-party/in-class, not as a claim of uniform category leadership, and note that GLM's own evaluation is only partially reproducible (the harness and data are not fully open).