Is GLM-5.1 open source?
Yes, with a practical qualification: Z.AI’s official Hugging Face repository labels GLM-5.1 as MIT-licensed and lists downloadable model-weight shards. The repository also provides usage guidance for Transformers, vLLM, and SGLang. These facts support a self-hosting route, but they do not mean that every related piece of code or every way of accessing the model has the same license or availability. (Official model repository)
A SourceForge mirror lists Apache License 2.0, but that is not the official model repository. In a Hugging Face discussion, a Z.AI representative distinguishes the model’s MIT license from Apache 2.0 licenses used for Z.AI code repositories. For commercial or other compliance-sensitive use, review the actual LICENSE file and notices that apply to the specific artifacts you use; the license label alone does not establish every applicable condition. (Z.AI license clarification)
Open-source status is not the same as API access
A license, downloadable weights, a hosted API, and availability in a particular model listing are separate questions:
- License: The official model repository labels GLM-5.1 MIT-licensed. Check the applicable license text and notices for your intended use; the label is not a substitute for reviewing those terms. (Official model repository)
- Weights and local runtimes: The repository lists
model-00001-of-00282.safetensorsand additional weight shards, and includes guidance for Transformers, vLLM, and SGLang. That establishes listed artifacts and documented runtime examples—not whether the model fits your hardware or production environment. (Repository files and guidance) - Hosted API:
Z.AI’s developer pricing page lists GLM-5.1. The page labels its text-model prices per 1 million tokens: US$1.40 input, US$0.26 cached input, and US$4.40 output. It lists cached-input storage as “Limited-time Free.” These are the rates shown in the documentation on November 9, 2026; confirm current pricing and the access terms for your account before budgeting or deployment. (Z.AIpricing) - Other providers and listings: Availability through one API or model catalog does not establish availability through another. Confirm the model identifier, account access, regional availability, limits, and supported features for the channel you intend to use. The sources cited here do not establish that TokenForge Cloud offers GLM-5.1, so verify its own model catalog and API documentation before planning an integration.
What the coding benchmarks suggest
Ollama’s GLM-5.1 listing reports the following GLM-5.1 versus GLM-5 results:
| Benchmark as labeled in the listing | GLM-5.1 | GLM-5 |
|---|---|---|
| SWE-Bench Pro | 58.4 | 55.1 |
| NL2Repo | 42.7 | 35.9 |
| Terminal-Bench 2.0 (Terminus-2) | 63.5 | 56.2 |
| Terminal-Bench 2.0 (“Best self-reported”) | 66.5 (Claude Code) | 56.2 (Claude Code) |
The captured table gives no units or methodology, so the first three rows are reproduced without adding percent signs. Keep the last row’s “Best self-reported” label: it is a separate comparison from the Terminus-2 row. These reported benchmark figures are not independent verification of performance on your codebase, API configuration, or production workload. (Ollama model listing)
The listing describes GLM-5.1 as a model for agentic engineering and says it can sustain work through repeated iterations and tool calls. Treat that as a capability claim from the listing, not a guarantee of latency, reliability, or results in a particular deployment.
Which route should you choose?
- Consider self-hosting if local inference fits your requirements and you can validate the license terms, obtain the listed artifacts, and meet the hardware and operations needs. The repository documents runtime examples, but the sources here do not establish your hardware requirements.
- Consider
Z.AI’s hosted API if managed inference is a better fit. Its pricing page lists GLM-5.1 and token rates; confirm account access, supported features, and current terms for your deployment. - Evaluate your own workload before committing. Compare task success, output quality, latency, total cost, and failure handling on the channel you intend to use. Public benchmark scores can help you shortlist a model, but do not replace that evaluation.
Bottom line: Z.AI’s official model repository labels GLM-5.1 MIT-licensed, lists weight shards, and provides runtime guidance—so it has a documented self-hosting path. Z.AI’s pricing documentation also lists GLM-5.1 for hosted access. Review the applicable license text and verify the access and production requirements of your chosen channel before relying on either route.
Sources
zai-org/GLM-5.1 — Hugging Face repository
GLM-5.1 license discussion — Hugging Face