Chinese AI company Z.ai revealed on August 26, 2026, that "Ox Alpha"—a model it had been offering anonymously on OpenCode and OpenRouter—was actually its new model, GLM-5.3-Flash. The model weights have already been released under an MIT license. According to Z.ai, all traffic during the anonymous testing period was processed on Chinese-made AI chips. The conversation has shifted from guessing the model's identity to how it was designed to run a massive multimodal model at low cost.
The Anonymous Testing and Official Release of Ox Alpha
Ox Alpha first appeared on August 20. OpenCode advertised a 1-million-token context window and multimodal input, offering it free for one week. The same anonymous model also appeared on OpenRouter, where developers used it for coding and long-running agentic tasks without knowing which company was behind it.
Six days later, Z.ai confirmed the connection to Ox Alpha in its announcement of GLM-5.3-Flash. The company said its goal was to gather user feedback anonymously before the official launch. Z.ai claims it became the most popular model within the week, though it did not disclose the reference population for that ranking or the volume processed. There is no external documentation that allows the usage figures from the anonymous test to be independently verified.
Along with the announcement, Z.ai released the model weights. Both FP8 and BF16 versions are available on Hugging Face, and the model can be self-hosted using SGLang or vLLM. Instructions for TokenSpeed and KTransformers were also provided—making it an open-weight model developers can deploy in their own environments.
Only 18 Billion of 320 Billion Parameters Active at a Time
GLM-5.3-Flash is a Mixture of Experts (MoE) model with 320 billion total parameters. Because it selects only the necessary expert components for each input token, just 18 billion parameters are actually active during inference. The design keeps the overall model large while limiting the computation required to process each token.
| Spec | GLM-5.3-Flash | GLM-4.5 Series | Change |
|---|---|---|---|
| Total Parameters | 320B | 355B | ~10% decrease |
| Active Parameters | 18B | 32B | ~44% decrease |
| Number of Layers | 45 | 92 | ~51% decrease |
While the total parameter count is close to that of the GLM-4.5 series, the number of parameters actually engaged and the layer count have both been roughly cut in half. Z.ai's basis for calling this model "Flash" lies less in shrinking the overall model and more in narrowing the computational path used during each inference. The company also says it used a new 30-trillion-token multimodal pretraining corpus, marking the first time images have been integrated into the foundation model within the GLM-5 series.
For connecting layers, the model adopts Manifold-Constrained Hyper-Connections (mHC). mHC projects the mixture of residual connections onto a specific manifold, preserving the properties of an identity mapping. The original paper proposed this technique as a way to curb the training instability that Hyper-Connections can cause, making it easier to scale up models. Z.ai pursued both directions simultaneously: growing the model while reducing computation at inference time.
Two Attention Mechanisms Supporting a 1-Million-Token Context
With long contexts, referencing every past token in detail each time causes computation and the KV cache to balloon. GLM-5.3-Flash uses linear attention to hold nearby dependencies as a state, while sparse attention retrieves relevant information from farther back. By splitting these two roles, the model avoids the burden of storing the entirety of long-text memory at the same precision and speed throughout.
Additionally, a mechanism called "IndexPool" compresses four key vectors used for retrieval into a single weighted vector. According to Z.ai's comparison, this reduces per-token attention computation to one-third and cuts the average KV cache per layer to roughly one-4.4th, compared to GLM-5.3. At the million-token scale, memory capacity and bandwidth tend to become bottlenecks before raw compute speed does. IndexPool addresses this burden directly.
However, the 3x and 4.4x figures do not represent overall response speed for the whole model. Z.ai normalized these numbers using per-head, per-layer computation and average BF16 KV cache to compare models of different scales. Actual speed will vary depending on output length, concurrent connections, and the inference engine's implementation.
Tens of Thousands of Chinese-Made AI Chips Bundled via EPD
Z.ai announced that all traffic during the anonymous testing period was processed using Chinese-made AI chips. The company says its cluster featured high-bandwidth interconnects and ran tens of thousands of Chinese-made accelerators. This scale of operation, even more than the model's identity reveal, is what defines this announcement.
Individual chips face constraints in memory capacity and bandwidth. Z.ai built a dedicated inference engine on top of SGLang, using intra-node tensor parallelism for the linear attention and output layers. W8A8 quantization is applied to computations. For the KV cache, the company uses INT8, FP8, and BF16 selectively—separating where precision is preserved from where capacity is reduced.
Across the cluster, the Encode stage (which converts inputs such as images), the Prefill stage (which processes input in bulk), and the Decode stage (which generates tokens one at a time) were split into separate worker pools. This is the Encode–Prefill–Decode (EPD) disaggregation architecture. Since each stage requires different amounts of compute and memory, this approach makes it easier to scale up only the congested stage rather than packing everything evenly onto the same servers.
Z.ai claims this improved end-to-end serving performance threefold compared to its initial configuration on the same hardware. The company also states that hardware efficiency and per-token cost are now on par with major NVIDIA GPUs. However, it did not disclose which GPU models, power consumption figures, or load conditions were used in the comparison. For now, this remains a self-reported evaluation by Z.ai.
The Meaning and Limits of the "100 Trillion Tokens a Day" Claim
The figure of "100 trillion tokens per day" that circulated around Ox Alpha is not the amount Z.ai actually processed. It was the daily service-wide capacity figure that OpenCode advertised when announcing the free trial—not a per-user allocation of 100 trillion tokens.
How this figure was counted has also not been disclosed. Are input and output tokens counted the same way? Whether cache tokens or internal reasoning tokens are included would significantly change the computational resources required. Z.ai's official announcement did not address this figure at all, stating only that "traffic from the past week was processed on a large-scale cluster of Chinese-made AI chips." Therefore, it cannot be said that 100 trillion tokens were actually processed daily on Chinese-made chips.
The same caution applies to the chips themselves. What Z.ai disclosed was limited to the category of "Chinese-made AI chips" or "accelerators developed domestically in China"—the manufacturer name and specific model were withheld. There is no public documentation confirming these are Huawei Ascend chips.
Some externally verifiable data on performance and pricing has begun to emerge. Artificial Analysis recorded a score of 57 on its own Intelligence Index and measured an output speed of 48.7 tokens per second via Z.ai's API. OpenRouter's standard pricing is $0.15 per million input tokens and $0.50 for output, with a 50% discount displayed at the time of publication. Z.ai states that it outperforms GLM-5.2 across multiple benchmarks and real-world tasks while cutting the price to one-tenth. However, the company has not disclosed the breakdown of token composition or discount conditions, so it cannot be assumed that all processing is uniformly one-tenth the cost.
GLM-5.3-Flash combines two claims: that the model weights are open, and that Z.ai's own service can run it at scale using only Chinese-made AI chips. If the specific chip models, power consumption, latency at million-token usage, and sustained throughput under congestion are independently verified by third parties, the field will move to a stage where it becomes possible to compare just how far Chinese-made accelerators can carry commercial-scale inference for massive models.