
A version number is not a product. That is the practical lesson of the week in open-weight models, because while Tencent put the weights of its new flagship Hy4 on Hugging Face on 20 November, the Chinese model everyone in the inference circles had been pricing against for weeks, Z.ai's GLM-5.5, does not exist.
OrcaRouter, an inference routing platform, published a teardown on 20 November checking five surfaces: Z.ai's release notes (newest entry: GLM-5.3-Flash, dated 26 August), Z.ai's documentation (the model-guide path for 5.5 returns a 404), the zai-org Hugging Face organisation (newest text repos are the GLM-5.3 family, created 25 August), Artificial Analysis (index page for 5.5 is a 404), and Z.ai's own rate card (nothing priced above GLM-5.3). No model card, no identifier, no rate, no weights.
Why the rumor was reasonable, and still wrong
Z.ai has shipped a numbered GLM release roughly every two months for a year: GLM-5 in February, 5.1 in April, 5.2 in June, 5.3 in August. Extrapolating a 5.5 in October or November was a fair inference from the cadence. The cadence has since broken rather than confirmed itself: the gap since GLM-5.3 is now four months, longer than any gap in the sequence.
The other engine behind the rumor is GLM-5.3-Flash, a 320-billion-parameter vision model with 18 billion active that shares the flagship's version stem. Pages that treat the Flash sibling as evidence of an imminent 5.5 are reading a product line as a countdown.
What actually shipped: Hy4
Tencent's Hy4 has been generally available via API since 4 November and open-weight since 20 November, when the company uploaded Hy4, Hy4-FP8 and Hy4-base to GitHub and Hugging Face, with GGUF builds, vLLM and SGLang images, and a system card. The language backbone is 770 billion parameters total, 49 billion active per token, with a 1.23-billion-parameter ViT encoder giving it native image and video input and a 1M-token context window.
On Artificial Analysis's Intelligence Index, Hy4 scores 35 against GLM-5.3's 32 at maximum effort. Three points is inside the noise of a single evaluation run, so the two shipped models are roughly equivalent in aggregate. The differences live in the rest of the spec sheet:
- Price. Hy4 runs $0.834 per million input tokens and $2.501 per million output; GLM-5.3 is $1.40 and $4.40. Tencent is 40% cheaper on input despite activating nine billion more parameters per token.
- Cache. $0.042 per million cached tokens for Hy4 against $0.26 for GLM-5.3, a six-fold gap that matters in agent loops that re-read a fixed context. Cost per completed index task: $1.03 versus $5.01.
- Modality. Hy4 reads images and video; GLM-5.3's flagship is text-only, with vision in the smaller Flash sibling.
- Where Z.ai wins. GLM-5.3 publishes a 128K output ceiling against Hy4's 64K, and reaches first token in 75.5 seconds against Hy4's 115.9.
The license line nobody priced
The Hy4 preview that shipped in August was Apache 2.0. The November general-availability build and the open-weight upload ship under the Tencent Hy License 1.0, which appends a Section 10: a model-as-a-service business above $50 million in worldwide gross revenue over any twelve months, or a product above 100 million monthly active users, needs a separate written license. Internal use is carved out. "Open weights" in November does not mean the same terms as "open weights" in August, and teams building on the earlier license grant should read Section 10 before shipping.
GLM-5.3 also carries a custom Z.ai license rather than a permissive standard, so the practical difference between the two shipped models is small. The difference that matters is between a model with checkable artifacts and a version number with none. Last week Cursor told House committees it may phase out Chinese open-weight models, which makes the identity and license of whatever you download less academic, not more.
What to watch
The trigger for a real GLM-5.5 is specific: a documentation page that resolves, a rate row above GLM-5.3's, and a weight repository in the zai-org organisation. One is a weak signal. All three together is a launch. If you find a page pricing GLM-5.5 before that, the page is describing a version number.

Sources
- OrcaRouter: Tencent Hy4 vs GLM-5.5: One of These Models Does Not Exist
- Tencent AI on X: Hy4 open-weight announcement, 20 November 2026
- Neurohive: Hy4 by Tencent: Open Weights, Benchmarks and Pricing
- AIToolsRecap: Tencent Hunyuan Hy4: 770B With Vision and Computer Use
- Shattered.io: Tencent Hy4 Is a 770B Apache 2.0 Open Model
