GLM-5.5 said to be in final stage of post-training
Information has emerged that "GLM-5.5," the next-generation model from Chinese artificial intelligence (AI) company Beijing Zhipu Huazhang Technology (Zhipu, known overseas as Z.ai), has entered the final stage of development. LuminaBench, an X account that tracks AI model developments, posted on Oct. 31 (7:33 p.m. the same day Japan time) that GLM-5.5 is undergoing "post-training," which refines performance after pretraining. Three internally developed models with 32 billion, 122 billion and 256 billion parameters (a measure of an AI's size) will be merged into the next-generation base model, according to the post. According to overseas forums and others that reproduced it, the post cited as training priorities environments for quantitative financial analysis, isolated environments for programming and a mechanism in which AI improves finance-related code on its own. As of the night of Nov. 2 (Japan time), Zhipu had not announced GLM-5.5 and had not disclosed its size, release date or pricing.
Poster denies three separate releases, gives no evidence
After the post, the interpretation spread that GLM-5.5 would be released in three sizes. LuminaBench posted again at 10:36 p.m. the same day (Japan time), explaining that the three are internal models being refined for merging into the next base and will not be released separately. The poster called the source of the information a "commit" (a record of a software change), but did not say which repository. As of Nov. 1, no changes supporting GLM-5.5 had been found in major open-source inference software (Hugging Face's transformers and SGLang), according to a check by OmniaKey. The merging method and the final model's size are also unknown. On Oct. 21, there was also a rumor that the next base would have 2.5 trillion to 3.5 trillion parameters, which conflicts with this information.
BTW you're not getting separate 32B / 122B / 256B GLM 5.5 models These are internal models being post trained to merge into the next gen base model (GLM 5.5) A lot of people took what I reported on and spun it into 3 separate model releases when that's not what the commit says
August release forecast missed, 2 1/2 months since GLM-5.3
A forecast for GLM-5.5's release timing has already missed once. Reuters reported in late June that the next model, GLM-5.5, was expected to be released in August. A report by JPMorgan suggested its size would exceed 1 trillion parameters (according to Startup Fortune). Zhipu never confirmed this, and what it released on Aug. 14 was "GLM-5.3," which improved the same base model as GLM-5.2 through post-training alone.
As of the Oct. 31 post, 78 days had passed since GLM-5.3's release. In 2026, the intervals between flagship model releases were 70 days from the release of GLM-5.1's weights (April 7) to GLM-5.2 (June 16), and 59 days from GLM-5.2 to GLM-5.3.
- GLM-5 releasedOfficial
- GLM-5.2 released, with context extended to 1 million tokensOfficial
- Reuters reports GLM-5.5 expected to be released in AugustReport
- GLM-5.3, an improved version on the same base, releasedOfficial
- GLM-5.3 weights released (Aug. 28 China time)Official
- Rumor that the next base will have 2.5 trillion to 3.5 trillion parametersRumor
- Post says GLM-5.5 is in final stage of post-trainingRumor
GLM-5.3 trailed two U.S. companies on terminal tasks
GLM-5.3 is designed to activate only 40 billion of its 744 billion total parameters in each pass, and the amount of text it can read at once (context) is 1 million tokens. In Z.ai's August announcement, its accuracy on Terminal-Bench 3.0, which tests solving terminal tasks, rose from GLM-5.2's 4.6% to 28.3%, but fell short of OpenAI's GPT-5.6 Sol (34.6%) and Anthropic's Claude Fable 5 (33.7%). The weights were released on Aug. 28 (Aug. 29 Japan time), under a custom license requiring organizations with annual revenue of more than $10 billion to undergo a safety review by Z.ai. Developer API pricing is $1.40 for input and $4.40 for output per million tokens.
January-June revenue up about fivefold, but short of market expectations
Zhipu listed on the Hong Kong stock exchange on Jan. 8 (stock code 2513). The offering price was 116.20 Hong Kong dollars, and it raised $558 million.
Revenue for January-June 2026, announced on Aug. 31, was 953.89 million yuan, about five times the figure a year earlier (up 399.7%). Cloud-based delivery accounted for 825.18 million yuan, or 86.5% of the total. A year earlier, the share was 15.2%. The net loss narrowed to 2 billion yuan from 2.4 billion yuan a year earlier, but research and development spending rose 36.6% to 2.1 billion yuan.
Revenue fell short of the market forecast of 1.35 billion yuan, and as of the end of August the stock was about 60% below its June high of 2,980 Hong Kong dollars (according to AI Weekly). JPMorgan forecasts full-year 2026 revenue of about 5 billion yuan. The timing of the next base model's completion will be a gauge of revenue growth.
- Cloud-based
- Other
