On the evening of August 26, the true identity of the mysterious model “Ox Alpha,” which had been running anonymously on OpenRouter for a week, was officially revealed: it was GLM-5.3 Flash from China’s Zhipu AI. The model features a MoE architecture with 320B total parameters and 18B active parameters, supports a 1 million token context window and native multimodal capabilities, and has its weights open-sourced on Hugging Face under the MIT license. More notably, Zhipu AI stated that throughout the entire free preview period, GLM-5.3 Flash ran entirely on domestically produced Chinese AI chips, processing 100 trillion tokens per day.
From Ox Alpha to GLM-5.3 Flash
The story begins on August 20th. A man namedThe anonymous model “Ox Alpha” suddenly appeared on OpenRouter and OpenCode platforms., offering free service with a 1 million token context. This mysterious model quickly caused a sensation in the developer community because its performance is on par with top closed-source models.
From August 21 to 22, researchers used Java stack trace analysis, error code 1214 comparison, and a perfect 30/30 tokenizer match to identify Ox Alpha as Zhipu AI’s GLM-5.3 series model. The forensic findings from Oxford University researchers circulated widely in the community. On the morning of August 26, Bloomberg reported that Zhipu AI confirmed Ox Alpha as a new iteration of the GLM series and promised to release open weights that evening. Later that night, Z.ai officially announced that Ox Alpha is GLM-5.3-Flash, simultaneously open-sourcing MIT-licensed weights on Hugging Face.
Model Architecture: Hybrid Attention with Sparse Activation
GLM-5.3-Flash uses an entirely new foundation model training approach, with both the architecture and training recipe completely redesigned. For the first time in the GLM series, Zhipu has introduced a hybrid architecture that combines sparse attention with linear attention, paired with an IndexPool mechanism for long-context scenarios.
Technical highlights include:
- Manifold-Constrained Hyper-Connections（mHC）A new connection mechanism, paired with hybrid attention, greatly reduces long-context inference costs.
- 320B total parameters / 18B active parametersThe MoE architecture activates only 18B parameters per inference, balancing performance with efficiency.
- 1 million token contextNative support, not a plug-in extension.
- native multimodalSupports text, images, and videos alike, with visual capabilities integrated into the code loop (render → verify → correct).
- A 30-trillion-token multimodal pretraining corpus.
According to Zhipu’s data, compared to the non-Flash version of GLM-5.3, GLM-5.3-Flash reduces attention computation by a factor of 3.0 and shrinks KV cache by a factor of 4.4.
Benchmark scores: approaching top-tier closed-source models
Benchmark data released by Zhipu shows that GLM-5.3-Flash is on par with Claude Opus 4.8 and GPT-5.6 Terra across multiple key metrics:
- Terminal Bench 2.1: 84.3 points (Claude Opus 4.8: 85.0, GPT-5.6 Terra: 87.4)
- DeepSWE v1.163.4 points (previous GLM-5.2 was 46.2, an improvement of 17.2 points; Claude Opus 4.8 was 58.0)
- AutomationBench v1.0.6: 48.8 points (GLM-5.2 only 26.2, an improvement of 22.6 points; Claude Opus 4.8 is 41.0)
- GDPVal-AA v21773 points, surpassing Claude Opus 4.8 (1582), GPT-5.6 Terra (1571), and Gemini 3.7 Flash (1527), leading in enterprise-grade Agent workflow benchmarks.
- Toolathlon Verified78.4 points (GLM-5.2 scored 59.9, Claude Opus 4.8 scored 76.2)
- HLE w/ Tools: 55.3 points (Claude Opus 4.8: 57.9)
In Z.ai’s own Code Bench v1.0 test (using the Claude Code 2.1.207 testing framework), GLM-5.3-Flash scored 29.0 at maximum effort level, nearly matching Claude Opus 4.8’s 29.5.
However, GLM-5.3-Flash is not ahead across the board. GPT-5.6 Terra still maintains an edge on DeepSWE (69.6) and Terminal Bench (87.4), while Gemini 3.7 Flash ranks first on AutomationBench (52.3).
Pricing Strategy: The Price War of Flash-Class Models
GLM-5.3-Flash’s pricing is the most lethal part of this release wave:
- standard input: : $0.15 per million tokens (approximately NT$4.9)
- Standard output: $0.50 per million tokens (approximately NT$16.3)
- Cache input: Per million tokens US$0.03 (approximately NT$0.98)
The $0.03 cache input pricing is particularly noteworthy. For RAG (Retrieval-Augmented Generation) and Agent application scenarios that require large amounts of system prompts or repeated context, this price significantly reduces operational costs.
Zhipu is also offering GLM-5.3-Flash in the Coding Plan on Z.ai, giving subscribers three times the usage quota of GLM-5.3.
China’s Domestic Chip Full-Stack Implementation: A New Landscape for Infrastructure
In its launch blog post, Zhipu clearly stated: “All of this traffic is being served on domestically produced Chinese AI chips.” The weight of this statement should not be underestimated.
According to the official description, GLM-5.3-Flash’s infrastructure includes:
- Large-scale China-made accelerator cluster, paired with high-bandwidth interconnect
- Customized SGLang inference engine, partially self-built by the GLM-5.3 infrastructure Agent (the model optimizes its own service stack)
- EPD Decoupled Architecture: Encode / Prefill / Decode Worker Pools Operating Separately at Cluster Scale
- 3x improvement in end-to-end service performance on the same hardware, with per-token cost comparable to NVIDIA GPUs
For developers outside China, this means that MoE-architecture Flash models can be served at OpenRouter scale without relying on H100 clusters. This has important implications for predictions about declining inference costs (Is the thing Jensen Huang is worried about really coming?).
Differences from GLM-5.3
What needs to be clarified is that GLM-5.3-Flash and the one released on August 14 GLM-5.3 They are different products. GLM-5.3 is a base model with 743B parameters, focusing on coding and cybersecurity capabilities; its open-weight release requires security review. GLM-5.3-Flash, on the other hand, is the multimodal Flash version with 320B-A18B, released immediately under the MIT license.
Conclusion
The release of GLM-5.3 Flash marks a new phase: the “Ox Alpha playbook” of anonymous preview → community decoding → formal open-source, proving that open-weight models have already reached a level where they can compete with top-tier closed-source models in agent and code workflows. With the 320B-A18B MoE architecture, MIT license, input pricing of $0.15 per million tokens, and infrastructure running entirely on domestically produced Chinese chips, this model sends a clear signal across three dimensions: technical, commercial, and geopolitical.
For developers, the most practical impact is that in high-frequency, low-latency agent scenarios, there’s now a choice that’s extremely low-cost with performance close to frontier models. Ox Alpha’s free lunch is over, but the era of open-weight MoE models is just beginning.
Source: KOCPC Chinese