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] Z.ai says all of the model's production inference runs on a cluster of more than 100,000 Chinese-made AI accelerators, and that an agent powered by GLM-5.3 did much of the work of building and optimizing that serving system.^[17]

Key facts

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 it was responsible for Ox Alpha, now called GLM-5.3-Flash, and planned to price the model at US$0.15 per million input tokens and US$0.50 per million output tokens; Z.ai's shares rose as much as 12 percent that day.^[8]

Z.ai's named launch post and Hugging Face repository had appeared on August 26, the day before that report. 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.

Quantized local deployments

Unsloth published third-party GGUF quantizations derived from the Z.ai checkpoint. Its repository lists artifacts from 93.1 GB and 97.6 GB at nominal one-bit settings through 157 GB and 200 GB at nominal four-bit settings. The repository supplies commands for llama.cpp and other local runtimes, but the artifacts are conversions rather than official Z.ai checkpoints.^[15] File size and nominal bit width do not establish retained task accuracy. No defined retention benchmark was published.^[15]

Lucas Fulks reported a separate NVFP4 deployment across four NVIDIA DGX Spark systems. For that one configuration, he separately reported "750K context verified" and a "real 250K+-token prefill." He also reported decode rates of 48 tokens per second on coding tasks and 39 on prose, and aggregate throughput of 123 tokens per second with six concurrent requests.^[16] These are operator-reported observations from one system and workload mix, not a controlled comparison or a general performance specification. They also do not prove equivalent behavior at the checkpoint maximum of 1,048,576 tokens.

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]

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, Z.ai reported 29.0 for GLM-5.3-Flash at maximum effort and 29.5 for Claude Opus 4.8. Z.ai's documentation says the benchmark was run through Claude Code 2.1.207; the launch post originally said the same but later removed the version note.^[1][2]

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 infrastructure

Chinese accelerator cluster

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]

In a September 17, 2026 engineering post, Z.ai said it had built a complete production inference service from scratch on a cluster of more than 100,000 Chinese-made AI accelerators and that all production inference for GLM-5.3-Flash runs on that system.^[17] Neither post names the accelerator vendor or chip model. Z.ai's public descriptions of the cluster's size have changed:

Z.ai did not explain the edit or say whether the later figure reflects growth of the cluster or a different way of counting. The edit was also not carried through everywhere: Z.ai's developer documentation page for the model, last modified September 18, 2026, still describes serving "across tens of thousands of domestically developed accelerators."^[2]

Serving stack

Z.ai says the accelerators' limited memory capacity and bandwidth were the main constraint, especially with a one-million-token context window, a new model architecture, and multimodal requests. Its answer was a set of memory optimizations that trade extra computation for bandwidth and extra communication for device memory.^[1][17] The stack combines:

• intra-node tensor parallelism for the linear-attention layers and the LM head;
• ReplaySSM, which Z.ai describes as trading compute for memory;
• W8A8 quantization;
• mixed-precision cache quantization in INT8, FP8, and BF16;
• Layer Split;
• an Encode-Prefill-Decode (EPD) disaggregated architecture that schedules and scales multimodal encoding, prefill, and decoding as separate worker pools.^[1][17]

Z.ai reports that these changes raised end-to-end serving throughput about threefold over its initial baseline on the same hardware, and that the service went from initial model adaptation to production readiness in less than two weeks. It also says hardware utilization and per-token cost reached levels comparable to mainstream NVIDIA GPUs.^[1][17][18] The company published a throughput-trajectory chart but no baseline logs, cost data, or NVIDIA comparison methodology, so these remain vendor statements.

The GLM-5.3 Infra Agent and "dense feedback"

According to the September post, much of this engineering was carried out by an "Infra Agent" powered by GLM-5.3, not by the served GLM-5.3-Flash model. Engineers defined objectives and system boundaries and reviewed critical changes involving system architecture, asynchronous concurrency, and production risk. The agent handled analysis, hypotheses, and code changes, and an experimental environment returned the results.^[17]

Z.ai argues that the agent's effectiveness depended less on code generation than on feedback it could trace to a specific cause. An end-to-end result such as a throughput drop does not say which layer is responsible, so Z.ai wired correctness tests, runtime logs, execution traces, runtime events, microbenchmarks, and end-to-end metrics into the agent's loop. It calls this approach "dense feedback" and lists three requirements: feedback should be local (tied to specific launch parameters, code changes, kernels, input conditions, threads, or code paths), cheap and quick to obtain, and objectively verifiable against reference implementations or controlled experiments.^[17] The post gives three worked examples:

The GitHub record for PR #1180 shows it was merged on August 27, 2026. Its description applies TF32x3 precision to the state-update and transformation-merging kernels to avoid precision loss at long context.^[19] In the decode-kernel case, Z.ai notes that introducing ReplaySSM first made the kernel slower before the later optimizations.^[17]

Ox Alpha usage claim

The same post repeats Z.ai's account of the Ox Alpha preview on OpenCode and OpenRouter. It says that within a week of launch the model became the most-used model on both platforms and processed more than 62 trillion tokens in six days.^[17] The launch post had described it more loosely as "the most popular model of the week."^[1] Z.ai gave no split between the two platforms. OpenRouter's own published counts cover only its route; they are documented on the Ox Alpha page.

Assessment

The helper was GLM-5.3, not the served GLM-5.3-Flash model, even though both Z.ai posts describe the work as the model helping to optimize "its own" inference system. Both posts also describe engineers remaining in the development loop. Neither says that the agent modified model weights or performed self-training. The September post titles itself "Toward Recursive Self-Improvement" and presents the work as an early form of it, but it also states that Z.ai has "not yet reached recursive self-improvement" and that choosing objectives, setting boundaries, and assessing risk remain human responsibilities.^[17]

The evidence Z.ai published consists of the three case studies, four figures, and one upstream code change. The post does not quantify how much of the work the agent performed relative to engineers, and no independent audit accompanies it. The account therefore supports model-assisted infrastructure engineering. It concerns inference and does not establish that GLM-5.3-Flash was trained on the same chips.^[1][17]

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.

An outside cyber and safeguard evaluation that included Flash appeared later, from a competing developer. In a September 29, 2026 post focused on GLM-5.3, Anthropic's Frontier Red Team reported that one of its researchers used GLM-5.3-Flash, in an isolated test environment, to build an exploit for a vulnerability that had already been publicly disclosed. Anthropic also produced abliterated copies of both models; its figures show both rarely refusing clearly harmful requests on public refusal benchmarks after abliteration, while their GPQA Diamond and CyberGym scores stayed largely intact. The GLM-5.3 article covers the full analysis.^[21]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26Z.ai. "GLM-5.3-Flash: Frontier Intelligence, Flash Cost." August 26, 2026. z.ai/...glm-5.3-flash
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Z.ai Developer Documentation. "GLM-5.3-Flash." Accessed August 27, 2026. docs.z.ai/...glm-5.3-flash
3. ^1 ^2Z.ai Developer Documentation. "Pricing." Accessed August 27, 2026. docs.z.ai/...pricing
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Z.ai. "zai-org/GLM-5.3-Flash." Hugging Face model card and weights, accessed August 27, 2026. huggingface.co/...GLM-5.3-Flash
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Z.ai. "GLM-5.3-Flash config.json." Released checkpoint configuration, accessed August 27, 2026. huggingface.co/...config.json
6. ^1 ^2 ^3 ^4 ^5OpenRouter. "Ox Alpha." Historical preview page, accessed August 27, 2026. openrouter.ai/...ox-alpha
7. ^1 ^2 ^3 ^4OpenRouter. "Z.ai: GLM 5.3 Flash" and Models API. Accessed August 27, 2026. openrouter.ai/...glm-5.3-flash and openrouter.ai/...models
8. ^1 ^2 ^3 ^4Luz 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. ^1 ^2 ^3Artificial 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
15. ^1 ^2Unsloth AI. "unsloth/GLM-5.3-Flash-GGUF." Hugging Face model card, quantized artifacts, and local-runtime instructions, accessed August 28, 2026. huggingface.co/...GLM-5.3-Flash-GGUF
16. ^Lucas Fulks. "Final update (for now): GLM-5.3-Flash NVFP4 on 4 Sparks." X, August 27, 2026. x.com/...2093033043497787652
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Z.ai. "Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure." September 17, 2026. z.ai/...glm-built-its-inference-infrastructure
18. ^Z.ai (@Zai_org). "We're sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash." X, September 17, 2026. x.com/...2100481236364079277
19. ^1 ^2fla-org. "[CP] use tf32x3 affine chain in kcp." Flash Linear Attention pull request #1180, merged August 27, 2026. github.com/...1180
20. ^1 ^2Internet Archive. Snapshots of the GLM-5.3-Flash launch post's content bundle, August 26, 2026 14:10 UTC (contains "tens of thousands of domestically developed accelerators") and August 27, 2026 10:37 UTC (phrase removed). web.archive.org/...glm-5.3-flash-BlVEnoXa.js and web.archive.org/...glm-5.3-flash-CS8pYy5F.js
21. ^Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao, and Tripp Gallagher. "GLM-5.3 and the spread of advanced cyber capabilities." Anthropic Frontier Red Team, September 29, 2026. anthropic.com/...ad-of-advanced-cyber-capabilities