GLM-5.3-Flash

GLM-5.3-Flash is an open-weight, natively multimodal mixture-of-experts large language model released by Z.ai on August 26, 2026. Z.ai labels it a 320-billion-parameter model with 18 billion parameters active for each token. It accepts text and visual inputs, produces text, and has a configured context window of 1,048,576 tokens. Z.ai published the weights under the MIT License.^[1][2][4][5]

The model was available before its named release as Ox Alpha, an anonymous preview on OpenCode and OpenRouter that began on August 20. Z.ai later identified GLM-5.3-Flash as the model behind that preview.^[1][6][8] It is distinct from GLM-5.3, Z.ai's larger, text-only flagship announced earlier in August. Flash starts from a newly trained base model and uses a hybrid architecture that combines linear attention with DeepSeek Sparse Attention.^[1][2][5]

Key facts

Scroll sideways for more →

Release and Ox Alpha preview

OpenRouter and OpenCode introduced Ox Alpha without naming its developer. OpenRouter's historical page recorded a 1,048,576-token route that accepted text, images, and video and returned text. It described the preview as a reasoning model for coding and sustained agent work.^[6] Bloomberg News reported on August 27 that Z.ai had confirmed Ox Alpha was a new GLM-series model and intended to publish its weights that night.^[8]

Z.ai's named launch and Hugging Face repository followed on August 26. OpenRouter subsequently changed the Ox Alpha page to identify ZAI GLM-5.3-Flash as the revealed model, and it added z-ai/glm-5.3-flash as the production listing.^[6][7] Independent launch coverage also confirmed the release, parameter counts, multimodal interface, and public weights.^[8][10]

The identity announcement resolved the developer question but did not turn every preview-era observation into a model specification. The free preview price, temporary route identifiers, traffic totals, and route-specific retention terms described the Ox Alpha distribution period. They do not define the permanent price, architecture, or data policy of every GLM-5.3-Flash deployment.^[6][7]

Architecture

Model size and expert routing

The released configuration has 45 language layers with a hidden size of 4,096. The first three feed-forward layers are dense. Each of the other 42 layers uses a sparse MoE block with 288 routed experts and one shared expert. The router selects eight routed experts for each token. Z.ai's 18-billion active-parameter figure refers to the amount of the 320-billion-parameter model used for one token rather than the checkpoint's total capacity.^[1][5]

Hybrid attention schedule

The attention schedule repeats three Kimi Delta Attention layers followed by one sparse-attention layer through most of the network. In total, the configuration contains 34 linear-attention layers and 11 DeepSeek Sparse Attention layers.^[5] The Kimi Linear paper describes Kimi Delta Attention as a recurrent linear-attention mechanism with fine-grained gating, while the DeepSeek-V3.2 report describes DSA as an indexer that selects a limited set of earlier tokens for the more expensive attention calculation.^[11][12] Those papers document the component families, not GLM-5.3-Flash's overall performance.

Z.ai says the linear layers model local dependencies through a recurrent state and the sparse layers retrieve globally relevant context. Its IndexPool mechanism compresses four sparse-indexer key vectors into one before selection. The released configuration corroborates a pool size of four, a sparse top-k setting of 2,048 positions, and enabled index-key compression.^[1][5]

The model also uses Manifold-Constrained Hyper-Connections, or mHC, with four residual streams. The original mHC paper constrains the matrices that mix residual streams to preserve an identity-mapping property intended to improve training stability at scale.^[13] Z.ai attributes part of GLM-5.3-Flash's scaling efficiency to mHC, but it has not published an ablation that isolates mHC's contribution in this model.^[1][5]

Z.ai calculates that the hybrid design uses 3.01 times less attention computation and 4.44 times less key-value-cache memory than GLM-5.3 under its comparison method.^[1][2] These are vendor calculations. They are not measurements of complete application latency, memory use, or cost on every serving system.

Vision stack and training disclosure

The checkpoint includes a 24-layer vision encoder with a hidden size of 1,024, a configured image size of 448 pixels, 14-pixel patches, and a temporal patch size of two. The configuration contains separate image and video token identifiers.^[5] These fields establish a native visual path in the released artifact. They do not, by themselves, measure accuracy on images or video.

Z.ai says it trained the new base model on a 30-trillion-token multimodal corpus.^[1][4] The company has not published the corpus composition, data-provenance breakdown, deduplication method, training-compute total, or training duration. The model card points readers to the February 2026 GLM-5 technical report, which predates GLM-5.3-Flash and describes another model architecture.^[4][14] It therefore provides family background rather than a technical report for Flash.

Access and deployment

The direct Z.ai API uses the model code glm-5.3-flash. Z.ai documents required thinking, with no option to disable it, along with streaming, function calling, context caching, and structured output. The same documentation lists support for images, video, and files in addition to text.^[2]

Z.ai and the model card provide local-serving paths for SGLang, vLLM, TokenSpeed, and KTransformers.^[2][4] Public weights make self-hosting possible, but the checkpoint's size still requires substantial storage and accelerator memory. A small active-parameter count reduces computation per token; it does not reduce the storage needed for all expert weights to 18 billion parameters.

Context limits can differ between hosted routes. The checkpoint and Z.ai documentation specify 1,048,576 tokens. On August 27, OpenRouter's aggregate record listed 1,310,720 tokens while its top-provider field listed a 1,048,576-token context and a 131,072-token completion limit.^[5][7] Applications that depend on the full window must check the selected provider rather than treating an aggregator's largest route as a universal limit.

Pricing

Z.ai's metered list prices at release were US$0.15 per million input tokens, US$0.03 per million cached-input tokens, and US$0.50 per million output tokens. A 50 percent launch promotion reduced those prices to US$0.075, US$0.015, and US$0.25 through September 9, 2026 at 16:00 UTC.^[3]

Scroll sideways for more →

The launch rates were temporary and should not be used for estimates after the promotion ends. Z.ai also made the model available through the GLM Coding Plan and advertised three times the plan quota of GLM-5.3. Plan credits and API token prices are separate billing systems.^[2][3]

Evaluation

Vendor results

Z.ai reported a broad set of coding, agent, and professional-work evaluations. Among the published comparisons, GLM-5.3-Flash scored 63.4 on DeepSWE v1.1 versus 46.2 for GLM-5.2, and 48.8 on AutomationBench versus 26.2. On Z.ai Code Bench v1.0, a private evaluation run through Claude Code 2.1.207, Z.ai reported 29.0 for GLM-5.3-Flash at maximum effort and 29.5 for Claude Opus 4.8.^[1][4]

These are vendor-reported benchmark results. Z.ai disclosed different contexts, token limits, timeouts, sampling settings, and graders for different rows. The private Z.ai Code Bench task set and complete artifacts were not released, so its comparison cannot be reproduced from the launch post alone.^[1]

Independent measurement

Artificial Analysis tested the model through Z.ai's API. Its August 27 profile recorded a score of 57 on Artificial Analysis Intelligence Index v4.1.1, a median output speed of 50.2 tokens per second, and 1.47 seconds to first token. The index run produced 150 million output tokens and cost US$138.02 when calculated at the model's list prices.^[9]

Artificial Analysis classified the model as open weight, reasoning-enabled, and capable of text and image input. That evaluation did not independently test the anonymous Chinese-accelerator cluster, self-hosted deployments, video understanding, or reliable retrieval across the full one-million-token window.^[9]

Inference on Chinese accelerators

Z.ai says all GLM-5.3-Flash traffic during the anonymous preview was served on a large cluster of Chinese AI chips. The launch post describes an inference engine built on SGLang, with separate pools for multimodal encoding, prompt prefill, and token-by-token decoding. It also names W8A8 quantization, mixed INT8, FP8, and BF16 cache formats, ReplaySSM, and Layer Split as parts of the serving stack.^[1]

According to Z.ai, this cluster used tens of thousands of domestically developed accelerators and reached three times the end-to-end performance of the company's initial baseline on the same hardware.^[1] Z.ai did not name the accelerator vendor or model, publish the baseline logs, or provide an independent audit. The statement concerns inference during the preview. It does not establish that GLM-5.3-Flash was trained on those chips.

Documentation and evidence limits

The released weights, model card, configuration, API documentation, and price table make GLM-5.3-Flash substantially more reproducible than the Ox Alpha preview. The MIT license allows commercial use and redistribution subject to the license terms.^[4] The release is still open weight rather than a complete open-science package: Z.ai has not published the training corpus, full training code, Flash-specific technical report, or training infrastructure record.^[1][4][5]

No Flash-specific system card, public safety-evaluation report, red-team protocol, or documented knowledge cutoff accompanied the release.^[1][2][4] This absence does not show that no internal testing occurred, but it limits outside assessment. Likewise, a one-million-token capacity is an interface limit, not proof that the model uses every part of a long prompt with equal reliability.

References

1. ^Z.ai. "GLM-5.3-Flash: Frontier Intelligence, Flash Cost." August 26, 2026. z.ai/...glm-5.3-flash
2. ^Z.ai Developer Documentation. "GLM-5.3-Flash." Accessed August 27, 2026. docs.z.ai/...glm-5.3-flash
3. ^Z.ai Developer Documentation. "Pricing." Accessed August 27, 2026. docs.z.ai/...pricing
4. ^Z.ai. "zai-org/GLM-5.3-Flash." Hugging Face model card and weights, accessed August 27, 2026. huggingface.co/...GLM-5.3-Flash
5. ^Z.ai. "GLM-5.3-Flash config.json." Released checkpoint configuration, accessed August 27, 2026. huggingface.co/...config.json
6. ^OpenRouter. "Ox Alpha." Historical preview page, accessed August 27, 2026. openrouter.ai/...ox-alpha
7. ^OpenRouter. "Z.ai: GLM 5.3 Flash" and Models API. Accessed August 27, 2026. openrouter.ai/...glm-5.3-flash and openrouter.ai/...models
8. ^Luz Ding. "Z.ai's Shares Soar After Claiming Surprise AI Hit Ox Alpha (1)." Bloomberg News via Bloomberg Law, August 27, 2026. news.bloomberglaw.com/...odel-that-rivals-deepseek
9. ^Artificial Analysis. "GLM-5.3-Flash Intelligence, Performance & Price Analysis." Accessed August 27, 2026. artificialanalysis.ai/...glm-5-3-flash
10. ^Maria Deutscher. "Z.ai open-sources 'Ox Alpha' model as GLM-5.3-Flash." SiliconANGLE, August 26, 2026. siliconangle.com/...x-alpha-model-as-glm-5-3-flash
11. ^Kimi Team et al. "Kimi Linear: An Expressive, Efficient Attention Architecture." arXiv:2510.26692, 2025. arxiv.org/...2510.26692
12. ^DeepSeek-AI et al. "DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models." arXiv:2512.02556, 2025. arxiv.org/...2512.02556
13. ^Zhenda Xie et al. "mHC: Manifold-Constrained Hyper-Connections." arXiv:2512.24880, 2025. arxiv.org/...2512.24880
14. ^GLM-5 Team et al. "GLM-5: from Vibe Coding to Agentic Engineering." arXiv:2602.15763, February 2026. arxiv.org/...2602.15763