Introduction
Z.ai released GLM-5.3-Flash on August 26, 2026: a 320B-parameter Mixture-of-Experts model that activates 18B parameters per token, reads text, images and video natively, and carries a 1M-token context window. It is the first natively multimodal model in the GLM-5 series, and the first frontier open-weight release to combine sparse and linear attention in one stack.
The point of the release is cost. Against GLM-5.3, the flagship Z.ai shipped eight days earlier, the company reports 3.01x less attention computation and a 4.44x smaller KV cache — the two quantities that decide what a long context actually costs to serve. Weights are on Hugging Face under the MIT license.
The architecture: hybrid attention
Most long-context models pay for length twice — once in attention compute that grows with sequence length, once in a KV cache that grows linearly and dominates GPU memory during inference. GLM-5.3-Flash attacks both.
The language model is 45 layers deep, against 92 in GLM-5.3, and interleaves two layer types:
- KDA linear-attention layers, which trade the quadratic attention pattern for a linear-cost recurrence.
- NoPE sparse MLA layers, which keep full attention but restrict which tokens each query attends to, with no positional encoding.
On top of that sits an IndexPool mechanism that compresses key vectors by weighted pooling, and Manifold-Constrained Hyper-Connections (mHC) for scaling efficiency. Routing sends each token through 8 of 288 experts. Native weights are FP8, roughly 306 GiB — large to hold, but the slice activated per token is small.
The model was pre-trained on a 30T-token multimodal corpus.
A note on the context number
Z.ai's documentation and the release notes both give a 1M-token window (1,048,576). The Hugging Face model card states 300,000, and OpenRouter advertises 1,310,720 with a 48,000-token completion cap. The figures describe different things — what the architecture supports, what a given host has configured, what the card was written against — so check the limit of the endpoint you actually call rather than the headline.
Performance
Vendor-reported
Z.ai's own numbers put the model well ahead of GLM-5.2 and close to the current frontier on agentic coding. Blank cells mean the comparison was not published, not a zero:
| Benchmark | GLM-5.2 | GLM-5.3-Flash | Claude Opus 4.8 |
|---|---|---|---|
| DeepSWE v1.1 | 46.2 | 63.4 | — |
| AutomationBench | 26.2 | 48.8 | — |
| Terminal-Bench 2.1 | — | 84.3 | 85.0 |
| Z.ai Code Bench | — | 29.0 | 29.5 |
Two of these are vendor comparisons against the vendor's own older model, which is the easiest kind of win to arrange. The Opus 4.8 rows are the interesting ones, and the ones worth independent replication before you plan around them.
Independently measured
Artificial Analysis has run the model through its own harness, which makes it the first outside read available. It scores 57 on the Artificial Analysis Intelligence Index v4.1.1, ranking 3rd of 109 models — the index aggregates nine evaluations including Terminal-Bench 2.1, GPQA Diamond, Humanity's Last Exam, SciCode and AA-LCR.
The operational numbers are less flattering than the intelligence score, and matter more for an agent loop:
- Output speed: 50.2 tokens/second, against a 67 t/s median across measured models.
- Time to first token: 1.47 seconds — competitive for the class.
- Verbosity: 150M output tokens to complete the index, against a 100M median.
- Cost to run the full index: $138.02.
That combination is the real trade-off. The model is cheap per token and slow per token, and it emits about 50% more tokens than the median model to finish the same work. For batch jobs the token price dominates and this is an easy win; for latency-sensitive interactive use, the arithmetic is closer than the headline rate suggests.
What it is aimed at
Z.ai positions the multimodal side at vision-driven UI coding, computer use and GUI automation, Office deliverables (PPTX, PDF, DOCX, XLSX), 3D scene creation in Blender, game development with Godot, financial workflows and document research. The common thread is software that reads a screen and acts on it, rather than image captioning or visual Q&A.
Availability and pricing
- API model ID:
glm-5.3-flash - Weights:
zai-org/GLM-5.3-Flashon Hugging Face, MIT license - Price: $0.15 per 1M input tokens, $0.50 per 1M output tokens, with a 50% promotional discount through September 9, 2026
- GLM Coding Plan: fully available, with 3x the quota of GLM-5.3
- Prior life: it ran anonymously as "ox-alpha" on OpenRouter and OpenCode before the reveal
Conclusion
GLM-5.3-Flash is an efficiency release rather than a capability leap: roughly GLM-5.3-class quality made much cheaper to run at length. The claim worth watching is architectural — that interleaving linear and sparse attention cuts serving cost three- to fourfold without giving up precise long-context recall. Artificial Analysis confirms the intelligence, and also shows the cost is partly repaid in latency and token count. Before committing an agent pipeline, the two things to test on your own workload are whether the 1M-token window holds up under real retrieval, and whether the vendor's Terminal-Bench figure survives outside replication.
Sources
- GLM-5.3-Flash overview — Z.ai developer docs
- Z.ai release notes — dated entries for GLM-5.3 (August 18, 2026) and GLM-5.3-Flash (August 26, 2026)
- zai-org/GLM-5.3-Flash on Hugging Face
- GLM-5.3-Flash on Artificial Analysis — independent Intelligence Index v4.1.1 score, speed and cost measurements
- Z.ai launches GLM-5.3-Flash under MIT license — TestingCatalog
See the full GLM model family in our catalog, read our GLM-5 launch coverage, or look up Multimodal AI in the glossary.