Z.ai Confirms Ox Alpha Is GLM-5.3-Flash: A 320B MoE That Ran on Chinese Chips

Six days after a nameless coding model topped OpenRouter's global charts and drew forensic teams into a tokenizer investigation, Z.ai officially confirmed on Wednesday evening that "Ox Alpha" was GLM-5.3-Flash — a 320-billion-parameter mixture-of-experts model that the company says ran its entire anonymous preview on Chinese AI chips, not a single Nvidia GPU in the loop. The weights are now on Hugging Face under an MIT license, priced at $0.15 per million input tokens, and Z.ai's shares closed more than 12 percent higher in Hong Kong on Thursday.

The identity question was already largely settled before Z.ai spoke. By August 22, a community researcher had obtained a Java stack trace from the server by submitting a malformed request; the internal class path mapped directly to Z.ai's API infrastructure, making the attribution nearly certain at the code level. Bloomberg queried Zhipu on the morning of August 26 and received confirmation before the official weights dropped. What Wednesday's announcement added was architecture detail, pricing, and the first explicit confirmation that the free preview had served as a production-scale stress test for China's domestic chip ecosystem — a dimension the benchmarks alone could not have established.

Read more: Community forensics traced Ox Alpha to Zhipu before Z.ai confirmed

What Z.ai Actually Built: A New Architecture, Not a Fine-Tune

GLM-5.3-Flash is not a quantized or distilled version of the existing GLM-5.3 or GLM-5.2. Z.ai describes it as a model trained from a new base on a 30-trillion-token multimodal corpus — text, image, and video — with the architecture redesigned around a problem that had constrained the GLM family: the cost of serving a 1-million-token context window at production throughput.

The most consequential architectural change is a hybrid attention design, the first in the GLM series to combine two fundamentally different attention mechanisms in the same model. Standard transformer self-attention scales quadratically with sequence length: doubling context from 500,000 to 1 million tokens roughly quadruples the attention compute. GLM-5.3-Flash addresses this by giving 34 of its 45 layers a linear-attention mechanism called KDA (Kernel-based Dense Attention), which approximates softmax attention in O(n) compute. The remaining 11 layers use NoPE sparse MLA — a more computationally exact sparse mechanism that retrieves globally relevant context without rotary position embedding. The cheap layers handle local dependency; the expensive layers handle long-range retrieval. The result, according to Z.ai's own comparison with GLM-5.3, is roughly 3x less attention compute and a 4.4x smaller KV cache — figures that have not yet been independently audited but are specific enough to test. These architecture details are confirmed by the vLLM serving recipe, which documents the 45-layer structure, the KDA/MLA split, and the expert configuration independently.

Each token is routed through 8 of 288 expert feedforward subnetworks, with 3 additional dense leading layers. Total parameter count is 320 billion; parameters active per inference step are 18 billion — substantially fewer than the GLM-5.2 design, which carried approximately 32 billion active parameters. That reduction, combined with a 45-layer depth versus the 92-layer GLM-5.2, is how Z.ai achieves a list price roughly one-tenth of its predecessor on input tokens.

A mechanism called IndexPool compresses the indexer key vectors used by the sparse attention layers through weighted pooling — approximately 4-to-1 compression — which is the primary driver of the KV cache reduction at long context. Z.ai acknowledges a remaining weakness: GLM-5.3-Flash's KV cache is still slightly larger than the equivalent designs in Kimi K3 and DeepSeek V4 Flash, and the company says so explicitly in the model card, which is an unusual and useful form of transparency. The architecture also adopts Manifold-Constrained Hyper-Connections (mHC), a residual information mixing method that Z.ai credits with maintaining model quality across the shallower, leaner layer count.

The model is the first GLM-5 to be natively multimodal rather than text-only: image and video inputs are processed through the same model stack rather than a bolted-on vision adapter, and Z.ai trained it to inspect rendered interfaces and visual output as part of coding and agent workflows — useful for browser-based agents that need to assess what is actually on the screen.

The Chinese-Chip Story Is the Infrastructure Story

Z.ai's claim that the stealth week ran "entirely on Chinese AI chips" across approximately 100,000 accelerators is the part of the announcement that drew the most attention outside the developer community.

The chip vendor is not named in Z.ai's public materials. Based on Z.ai's training history — GLM-5 was trained on 100,000 Huawei Ascend 910B processors — the likely candidate is Huawei's Ascend family, but that cannot be stated as confirmed from company-level disclosures about this specific deployment. What the company does describe is the serving architecture: a custom SGLang-based engine with EPD (Encode-Prefill-Decode) disaggregation that separates multimodal encoding, prompt prefill, and token-by-token decoding into independently scheduled worker pools. Z.ai claims this yielded a 3x end-to-end serving improvement at cluster scale — also company-reported and not independently audited. The technical details of this serving approach are more interesting than the unnamed silicon: the architecture was designed specifically around the memory-capacity and memory-bandwidth constraints of the domestic accelerators, not against them.

What is independently verifiable is the output of that infrastructure. On OpenRouter alone, the model processed 23.2 trillion tokens across six days of stealth operation — roughly 2.3 times the next highest model on the platform during the same period. SemiAnalysis, the semiconductor research firm, flagged the disclosure that Z.ai had served the preview at scale on Chinese chips as the most significant element of the announcement. Separately, Artificial Analysis measured 48.7 output tokens per second and a 1.52-second time-to-first-token on Z.ai's API — independent performance data that confirms the model is serving at useful throughput on whatever hardware Z.ai is running.

The geopolitical dimension is specific: Z.ai and its subsidiaries have been on the US Commerce Department's Bureau of Industry and Security Entity List since January 2025, placed there on grounds that the company advances China's military modernization through AI. A company on the Entity List demonstrating that it can handle more global inference traffic than any prior OpenRouter launch — and do it without the Nvidia hardware that US export controls are designed to restrict — is a data point that export-control policy analysts will register. The question of whether Chinese chips can serve frontier models at global scale at competitive throughput just received a meaningful empirical answer, even if the hardware vendor remains unnamed.

Benchmarks: What Is and Is Not Independently Established

Z.ai reports scores on a set of agentic and coding benchmarks that show GLM-5.3-Flash near Claude Opus 4.8 on several metrics and ahead of GLM-5.2 by wide margins on others. Terminal-Bench 2.1, a benchmark for terminal-based agent tasks, puts GLM-5.3-Flash at 84.3 against Opus 4.8's 85.0 — a 0.7-point gap — though GPT-5.6 Terra scores 87.4 on the same benchmark. On DeepSWE v1.1, a software engineering benchmark, Z.ai reports 63.4, up from GLM-5.2's 46.2. AutomationBench, a workflow automation test, shows 48.8 against GLM-5.2's 26.2. These are all Z.ai-reported figures, running on company-selected harnesses and settings detailed in the model card footnotes.

The independent anchor comes from Artificial Analysis, which placed the model 10th on its Intelligence Index at a score of 57 — ahead of DeepSeek V4 Pro Max, according to independent measurement by that firm. That independent score, alongside the independently measured throughput figures, gives a more reliable baseline for production planning than benchmark tables compiled under proprietary evaluation conditions.

Vision remains a clear weakness. GLM-5.3-Flash trails Gemini 3.7 Flash on BabyVision and MVbench — acknowledged in Z.ai's own documentation. For applications that depend on precise image understanding rather than code inspection from screenshots, that gap matters.

Deployment Picture: API vs. Self-Hosting

The practical split for teams evaluating GLM-5.3-Flash is straightforward. The model's FP8 checkpoint weighs approximately 306 gigabytes before runtime and KV-cache overhead. The current vLLM path requires NVIDIA Hopper-generation GPUs or newer — an irony given Z.ai's Chinese-chip serving claim, but a reality for the open-source inference frameworks that most engineering teams use. Self-hosting at 4-GPU tensor parallelism requires at minimum a GB200 tray or an 8-GPU Hopper node with the prefill and decode pools split across the same machine. For organizations that do not have that hardware, the model is an API product.

Standard API pricing is $0.15 per million input tokens and $0.50 per million output tokens, with a 50 percent launch discount running through September 9 that cuts those rates to $0.075 and $0.25 respectively. Cached input runs at $0.03 per million. The GLM Coding Plan subscription tiers — Lite at $18 per month, Pro at $80, Max at $168 — each receive three times the usable token quota that GLM-5.3 carried, and multimodal capabilities appear inside ZCode through Browser Use and Computer Use features. KTransformers supports an alternative local-serving path for organizations with consumer-grade hardware at the cost of throughput. Full pricing details and Coding Plan information are available in Z.ai's published materials.

Open-Weight Positioning Against DeepSeek and Kimi

Three open-weight models now occupy the flash-tier multimodal coding space at meaningfully different architectural philosophies. DeepSeek V4 Flash uses its own Compressed Sparse Attention and Heavily Compressed Attention combination, with 49 billion active parameters from a 1.6-trillion-parameter base — a much larger total-parameter model with more active compute per token. Kimi K3, from Moonshot AI, is the most parameter-dense of the three at 2.8 trillion total parameters and 896 experts, making it also the most expensive to serve. GLM-5.3-Flash occupies a different efficiency profile: the smallest active-parameter footprint of the three at 18 billion, the shallowest layer count, and the hybrid linear/sparse attention design that Z.ai claims gives it the lowest per-token attention compute in the group — at the cost of a slightly larger KV cache than its two peers.

On Z.ai's own DeepSWE numbers, GLM-5.3-Flash scores 63.4 against DeepSeek V4 Flash-Vision-Exp's 59.3. On NL2Repo — a benchmark for translating natural language descriptions into repository-scale code — Claude Opus 4.8 scores 69.7 while GLM-5.3-Flash reaches 56.3, a gap that remains substantial at repo scale. The competitive claim is not that GLM-5.3-Flash beats everything but that it delivers near-frontier agentic performance at a price point that open-weight models have not reached before in a 320-billion-parameter package with native multimodality and an MIT license. Whether that claim holds under independent evaluation at scale is the question that comes next.

Read more: DeepSeek adds vision to V4 Flash, entering the multimodal agent space

The Stealth Launch as Product Strategy

The broader pattern matters. The OpenRouter stealth mechanism has now been used four times by Chinese labs in 2026, with each use refining what the format delivers: a week of production traffic, real developer feedback, and an adoption base before the named launch carries institutional reputational weight. GLM-5.3-Flash used that week more effectively than any prior stealth model — processing more traffic, attracting more users, and building a more legible community investigation into its identity. Z.ai's timing of the reveal, releasing weights the same evening Bloomberg confirmed the identity through traditional reporting, suggests the company was watching when the journalistic confirmation arrived and accelerated the official announcement accordingly.

The practical effect is a named product with a publicly documented architecture, priced API access, downloadable weights, and an established developer audience that already has opinions about the model's performance on their actual workloads. The free period is over. The question now is whether the throughput, price point, and MIT license are sufficient to convert the 300,000-plus developers who used Ox Alpha into paying Z.ai API customers — or whether they route to the weights directly and build self-hosted pipelines that Z.ai does not directly monetize. That conversion rate, not the benchmark table, is the real commercial test of whether a stealth week is worth 62 trillion tokens of free inference.