Zhipu has released its latest model, GLM-5.3-Flash, which was previously made available for free on a testing platform under the anonymous name "Ox Alpha," attracting over 50 trillion tokens in traffic within five days. The model utilizes over 100,000 domestic chips for inference services, achieving hardware efficiency and cost per token comparable to NVIDIA GPUs. In terms of performance, the model scores 57 on the AA Comprehensive Intelligence Index, matching Claude Opus 4.8. Pricing is set at 0.8 yuan per million tokens for input and 2.8 yuan per million tokens for output—significantly lower than competitors. The architecture features a hybrid design of sparse and linear attention, making it the first natively multimodal model in the GLM-5 series. Amid a widespread trend of price increases among domestic AI models, Zhipu is leveraging a low-price strategy to capture market share.Author and source: LatePost
On August 26, Zhipu officially confirmed: the anonymous model Ox Alpha, which sparked widespread discussion in the developer community (known in Chinese communities as "Niu Lai"), is none other than its newly released GLM-5.3-Flash. The model's codename was inspired by the recently released Chinese film "Niu Lai."
Before its official launch, the model was made available for free testing in an anonymous format on OpenRouter and OpenCode, accumulating over 50 trillion tokens in traffic within five days—setting new traffic growth records on both platforms—and its current usage has surpassed twice that of DeepSeek. However, what truly captured market attention was a subsequent disclosure by Zhipu: all this traffic was entirely supported by domestic chips.
Zhipu stated in its official technical documentation that over 100,000 domestic chips are deployed in a cluster for this model's inference service, "achieving hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs."
The semiconductor research firm SemiAnalysis immediately posted a comment on X: "All traffic is carried by domestically produced chips, with hardware efficiency and cost per token now comparable to NVIDIA GPUs. Following yesterday’s announcement of Jalapeño (OpenAI’s in-house inference chip), the CUDA moat is once again under scrutiny."
100,000 domestic GPUs: Zhipu describes the deployment details of this domestic chip cluster in its technical documentation.
The primary bottleneck of a single chip lies in memory capacity and bandwidth, making it especially challenging to support context lengths of up to 1 million tokens. To address this, Zhipu has built a dedicated inference engine on top of SGLang, incorporating technologies such as W8A8 quantization, mixed-cache quantization with INT8/FP8/BF16, and intra-node tensor parallelism. Additionally, it introduces a production-grade, separated Encode–Prefill–Decode (EPD) architecture, decoupling multimodal encoding, prompt prefilling, and token-by-token decoding into independently schedulable work pools.
Zhipu stated that end-to-end service performance has improved threefold compared to the initial baseline on the same hardware.
According to LatePost, the suppliers of these chips may include Huawei, Moore Threads, and Hygon. Zhipu has not commented on this, and the technical documentation does not specify the exact chip models.
SemiAnalysis specifically noted in its commentary that it was previously believed only top-tier frontier labs could achieve a computational scale of 100 trillion tokens per day, "yet here, 100 trillion tokens per day of free traffic are all running on domestic chips."
Model specifications: Comparable to DeepSeek, priced at 1/40th of Claude Opus. GLM-5.3-Flash has a total of 320 billion parameters and 18 billion activated parameters, representing approximately 40% of the parameter scale of the previous flagship model, GLM-5.3, and nearly matching DeepSeek V4 Flash.
In terms of performance, official evaluations by Zhipu show that GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index (AA Intelligence Index), surpassing the previous flagship GLM-5.2, matching Anthropic’s Claude Opus 4.8, and exceeding DeepSeek’s flagship model V4 Pro’s score of 53.
Pricing for GLM-5.3-Flash is RMB 0.8 per million tokens for input, RMB 2.8 per million tokens for output, and RMB 0.23 for cache hits—one-tenth the price of GLM-5.3. A 50% discount is available during the first two weeks after launch.
Zhipu stated that GLM-5.3-Flash is priced at one-tenth of GLM-5.3, and during the limited-time discount, it is one-twentieth of GLM-5.3 and one-fortieth of Claude Opus 4.8.
Compared to the revised pricing of DeepSeek V4 Flash (1.5 RMB per input, 4.5 RMB per output), GLM-5.3-Flash is cheaper in the vast majority of standard usage scenarios.
Architectural innovation: The number of activated parameters and layers is nearly halved. GLM-5.3-Flash employs a completely different architecture from GLM-5.3, which is the core reason it achieves higher performance at a lower cost.
Compared to GLM-4.5, GLM-5.3-Flash has a similar total parameter count (355B vs. 320B), but the activated parameters have decreased from 32B to 18B, and the number of layers has been reduced from 92 to 45, nearly halved.
Zhipu AI states that GLM-5.3-Flash is the first open-source frontier model to adopt a hybrid architecture combining sparse attention and linear attention. Compared to GLM-5.3, it reduces attention computation by 3.01 times and KV cache size by 4.44 times.
In addition, this model is the first natively multimodal model in the GLM-5 series, supporting image and video inputs, and marks the first new model with multimodal capabilities launched by Zhipu since it strategically focused on coding.
The release of GLM-5.3-Flash comes after a wave of price increases across China's AI model market.
Kimi K3's output price reaches up to 100 yuan per million tokens, more than three times higher than its predecessor K2.6; after a price increase in the first half of the year, Zhipu's output price has also reached 28 yuan; DeepSeek V4 Pro has risen from 6 yuan to 13.5 yuan, with peak-hour pricing at 27 yuan; DeepSeek V4 Flash's output price has increased from 2 yuan to 4.5 yuan, with peak-hour pricing at 9 yuan.
The direct consequence of the price increase is demand spillover. According to LatePost, after DeepSeek V4 Flash raised its prices, the number of API calls on the OpenCode platform dropped by half, creating new growth opportunities for domestic model providers.
The pricing strategy of GLM-5.3-Flash is precisely designed to address this gap.