Z.ai's GLM-5.5 is on the roadmap, but its sizes and specs are unconfirmed

Z.ai’s next major model, GLM-5.5, has been reported as a later milestone on the company’s roadmap, but the documents reviewed for this article contain no official launch date, architecture, parameter sizes or licence for it. Z.ai’s own public material still presents GLM-5.3 and GLM-5.3-Flash as its newest models, and those releases are the firmest guide to what GLM-5.5 is likely to build on.

Builders tracking the GLM family should treat any specific GLM-5.5 sizes or launch dates as unconfirmed until they appear in a Z.ai model card, repository or API listing. This article sets out what is documented, what is only reported, and where the gaps are.

Where GLM-5.5 stands: reported, not documented

The clearest summary of GLM-5.5’s status comes from CometAPI, an API reseller that published an explainer in August and updated it on 3 September 2026. It says Reuters described GLM-5.5 as a later roadmap milestone, but that the report “did not provide a confirmed launch commitment, architecture, parameter count, context limit, benchmark table, price, or access plan”.

According to the same explainer, the “5.5” name had appeared in Reuters reporting but not in an official Z.ai model card, developer endpoint, release note, Hugging Face repository or pricing table. It also states that no public evidence supported a specific trillion-parameter claim for GLM-5.5.

The text of Z.ai’s GitHub repository reviewed here is consistent with that picture: its headline lists GLM-5.3, GLM-5.2, GLM-5.1 and GLM-5, and directs developers to GLM-5.3 and GLM-5.3-Flash on the Z.ai API Platform. No GLM-5.5 entry appears in it.

The one directional signal CometAPI cites is from Z.ai’s CodeGeeX technical lead, who told Reuters that future models would target long-horizon tasks and “self-evolving autonomous agents”. CometAPI reads this as a sign that GLM-5.5 will be judged on reliable task completion, self-correction and sustained engineering work rather than a single headline parameter increase. That is CometAPI’s interpretation, not a Z.ai statement.

The GLM-5 line so far: what Z.ai has published

Each step in the GLM-5 series has been documented by Z.ai, which gives a concrete baseline for the next release.

Model	What Z.ai says changed
GLM-5	744B parameters (40B active), up from 355B (32B active) in GLM-4.5; pre-training data up from 23T to 28.5T tokens; adds DeepSeek Sparse Attention
GLM-5.1	Built to stay effective on agentic tasks over longer horizons, sustaining optimisation over hundreds of rounds and thousands of tool calls
GLM-5.2	A 1M-token context; multiple thinking effort levels; IndexShare architecture change
GLM-5.3	Same base model as GLM-5.2; all gains from post-training
GLM-5.3-Flash	Newly trained base model with hybrid sparse and linear attention, trained on a 30T-token multimodal corpus

Two terms matter here. A sparse mixture-of-experts (MoE) model holds many parameters but activates only a fraction for each token, which is why GLM-5 is described as 744B total but 40B active. CometAPI cites the GLM-5 technical report as describing 256 experts, eight routed experts per token and one shared expert. Because GLM-5.3 keeps the same base model, CometAPI treats that roughly 744B-total, 40B-active design as the closest published architectural reference for whatever comes next.

CometAPI also states that GLM-5.3 offers a 1M-token context window and a 128K maximum output. Z.ai’s repository text confirms the 1M-token context from GLM-5.2 onward but the reviewed excerpt does not state the 128K output limit.

Training pipeline clues in Z.ai’s own documents

Z.ai has not published a GLM-5.5 training recipe in the sources reviewed. Its recent releases do, however, describe the pipeline pieces it is investing in.

A new base model for GLM-5.3-Flash

The most significant change is that GLM-5.3-Flash “starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency”. Z.ai says it is the first GLM model to use a hybrid architecture combining sparse and linear attention, which it says sharply reduces long-context serving costs while preserving precise long-context capabilities. Attention is the mechanism a model uses to relate tokens to one another; sparse and linear variants are cheaper to compute over very long inputs than standard attention.

The Flash model also adopts Manifold-Constrained Hyper-Connections (mHC), which Z.ai says improves scaling efficiency, and it was pre-trained on Z.ai’s “latest 30T-token multimodal pre-training corpus”. Z.ai says these changes let GLM-5.3-Flash “deliver more intelligence with less compute”. A fresh base model, a new attention design and a larger multimodal corpus are the kind of foundation a next-generation release could reuse, though no source confirms that GLM-5.5 does.

Post-training and reinforcement learning

GLM-5.3 shows how far Z.ai is pushing post-training, the stage after pre-training in which a model is tuned with methods such as reinforcement learning (RL). Z.ai says every gain in GLM-5.3 over GLM-5.2 comes from post-training. For GLM-5, it built an asynchronous RL system called slime to raise training throughput and allow “more fine-grained post-training iterations”.

Efficiency work on long context

For GLM-5.2, Z.ai introduced IndexShare, which reuses the same indexer across every four sparse attention layers and, by Z.ai’s figures, cuts per-token FLOPs (floating-point operations, a measure of compute) by 2.9× at a 1M context length. It also says it improved the multi-token prediction (MTP) layer used for speculative decoding, raising acceptance length by up to 20%.

Multimodal agents

A separate technical report from Z.ai and Tsinghua University, posted on 29 April 2026, describes GLM-5V-Turbo, a model that treats vision as a core part of reasoning and tool use rather than an add-on. It introduces CogViT, a new vision encoder, and Multimodal Multi-Token Prediction, and describes joint reinforcement learning over more than 30 task categories spanning perception, reasoning and agentic capabilities. The report’s authors conclude that “agentic competence is often acquired more effectively through hierarchical optimization than through monolithic end-to-end training”.

Maker claims versus independent evidence

Almost all performance numbers for the GLM-5 series come from Z.ai itself. In Z.ai’s own figures, GLM-5.2 scores 81.0 against GLM-5.1’s 62.0 on Terminal-Bench 2.1 and 62.1 against 58.4 on SWE-bench Pro, and lands within a few points of Claude Opus 4.8 (85.0) on Terminal-Bench 2.1 while staying ahead of Gemini 3.1 Pro.

For GLM-5.3, Z.ai reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, and claims open-source state of the art on Terminal Bench 3.0 and Agents’ Last Exam. An in-house benchmark cannot be checked by outsiders, so the 50% figure is best read as directional.

For GLM-5V-Turbo, the technical report lists 62.3 on OSWorld and 75.7 on AndroidWorld for computer and phone interface tasks, and 94.8 on Design2Code, which it says outperforms Claude Opus 4.6.

The Hugging Face page for the original GLM-5 lists a GPQA Diamond score of 86 sourced to the model card and a SWE-bench Verified score of 72.8 attributed to the SWE-bench official evaluation, with both entries marked as not verified on the page. The same page shows GLM-5 served by Z.ai and third-party inference providers including Novita, Featherless AI and DeepInfra, and records more than 3.3 million all-time downloads. That is evidence of real adoption of the earlier model, not of GLM-5.5’s capabilities.

None of the sources contains an independent benchmark of GLM-5.5, which is expected given that no official release is documented.

What it means for builders

For teams already using GLM models, the practical baseline is GLM-5.3 and GLM-5.3-Flash, which Z.ai offers through its API platform and its z.ai chat site. Planning around specific GLM-5.5 sizes would be premature on the evidence available.

• Long-running agents. Z.ai’s stated direction, from GLM-5.1 onward, is models that keep improving over hundreds of rounds and thousands of tool calls. Evaluation harnesses that measure final task outcomes over long sessions will matter more than single-turn tests, a point CometAPI also makes.
• Serving cost. The hybrid attention in GLM-5.3-Flash and IndexShare in GLM-5.2 both target the cost of long contexts. If those carry forward, long-context workloads are where any savings would show up first.
• Open weights. CometAPI considers a continued open-weight release plausible because GLM-5.3 is positioned as open-weight, but says the GLM-5.5 licence remains unknown. Teams that depend on self-hosting should wait for the actual licence text. Licence terms have varied across recent large open releases, as with Baidu’s regionally restricted ERNIE base model.

Risks and open questions

Z.ai’s own GLM-5.3 notes flag a safety-relevant trend. The company says that as it scaled post-training, “cyber capability developed faster than we expected,” that GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and that it more than doubles GLM-5.2 on exploitation benchmarks. If GLM-5.5 continues the post-training push, how Z.ai handles offensive security capability in an openly distributed model is an open question the sources do not answer.

There is also a source-quality caveat. The main summary of GLM-5.5’s status comes from a reseller that says it expects to add GLM-5.5 quickly if the model is released, so its framing has a commercial interest. Its factual statements about what Reuters did and did not report are the useful part.

What to watch next

• An official Z.ai model card or repository entry naming GLM-5.5, with parameter counts, active parameters and context length. Until then, any reported sizes remain unconfirmed.
• Whether GLM-5.5 builds on the new GLM-5.3-Flash base with hybrid sparse and linear attention and the 30T-token corpus, or on the 744B GLM-5 base used through GLM-5.3.
• The licence, and whether weights appear on Hugging Face alongside API access.
• Independent results on public agentic and coding benchmarks, rather than Z.ai’s in-house Code Bench alone.
• Safety disclosures covering the cyber capabilities Z.ai has already said grew faster than expected in GLM-5.3.

Unicorn API will update this thread when Z.ai publishes official GLM-5.5 details.