GLM 5.3 Flash: Ox Alpha Stealth Model Revealed by ZAI
ZAI confirms Ox Alpha was GLM 5.3 Flash, a 320B MoE model given MIT weights after serving 44 trillion tokens in stealth testing.
What was Ox Alpha, and what did ZAI reveal?
Ox Alpha was an anonymous stealth model that appeared with a 1 million token context window, multimodal support, and free usage, prompting months of community fingerprinting. On August 26, ZAI (formerly Zhipu AI) confirmed it was GLM 5.3 Flash, the first natively multimodal model in the GLM 5 series. ZAI released full weights on Hugging Face under an MIT license, published technical specs, and disclosed that the entire stealth preview ran on Chinese AI chips rather than Nvidia hardware.
TL;DR
- ZAI confirmed that the mystery model Ox Alpha, tested anonymously for months, is GLM 5.3 Flash, ending a long round of community fingerprinting based on tokenizer and error-code matches.
- The model shipped as fully open weights on Hugging Face under an MIT license, with 320 billion total parameters and only 18 billion active per token in a mixture-of-experts design.
- The stealth run was massive, serving 44 trillion tokens to over 500,000 unique users across more than 13 million sessions, at one point accounting for over 10% of all usage on Open Code.
- API pricing landed at 15 cents per million input tokens and 50 cents per million output tokens, with cached input priced at 3 cents, making it one of the cheapest frontier-adjacent models available.
- Independent retesting on the official API scored lower than the stealth version on one-shot visual tasks, though agentic and reasoning benchmarks held steady at top marks.
- The architecture is built for local inference, since the low active-parameter count keeps compute costs down even though the total parameter count requires substantial memory.
- ZAI disclosed the stealth preview ran entirely on domestic Chinese AI chips using a custom SGLang-based inference engine, which the company says delivered roughly a 3x improvement in serving performance over prior setups.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How did the community figure out Ox Alpha’s identity before the reveal?
Before ZAI’s official statement, independent testers pieced together the model’s origin through technical fingerprinting rather than any leak. The tokenizer used by Ox Alpha matched GLM 5.3’s tokenizer exactly. Its video encoder behaved identically to GLM5V’s. Even the error codes returned by the API matched ZAI’s own systems. That combination of evidence pointed strongly at ZAI well before the official confirmation, though the working assumption was that it might be a smaller closed variant rather than something ZAI would eventually open-source. The final reveal split that guess: the model was indeed the smaller member of the GLM 5 family, but it was set free entirely rather than kept closed.
What are the specs of GLM 5.3 Flash?
GLM 5.3 Flash is a mixture-of-experts (MoE) model with 320 billion total parameters and 18 billion active parameters per token. That gap matters: total parameters determine memory footprint, while active parameters determine actual compute and inference speed. The model uses a hybrid attention architecture combining sparse and linear attention to cut the cost of serving long context, along with a technique ZAI calls Manifold-Constrained Hyper-Connections (mHC) aimed at improving scaling efficiency. It was trained on a 30-trillion-token multimodal pretraining corpus and supports a controllable reasoning effort parameter (low, high, or max) for balancing latency against depth of reasoning.
On ZAI’s own benchmark reporting, the model scores 84.3 on Terminal-Bench 2.1, close to Claude Opus 4.8’s 85. On DeepSWE it scores 63.4, up sharply from GLM 5.2’s 46.2, and it also shows a large jump on AutomationBench compared to its predecessor. The weights are published on Hugging Face as safetensors files (split across 62 shards) under an MIT license, meaning anyone can download, modify, fine-tune, or redistribute the model without restriction.
How does GLM 5.3 Flash compare against the stealth version?
Independent retesting through the official API, using an eight-question benchmark suite covering coding, visual one-shot generation, math reasoning, and an end-to-end fine-tuning pipeline, produced a total score of 63 out of 80 (78.75%). During the anonymous stealth period, the same suite scored 70 out of 80 (87.5%) under the Ox Alpha name, a gap of nearly nine percentage points.
The drop was concentrated entirely in visual one-shot tasks: an elevator simulation with boarding logic, a 3D contact lens case model, and an SVG illustration of a panda eating a burger. All three scored lower on the official release than they had during stealth testing. Meanwhile, tasks tied to agentic behavior and structured reasoning, including a hard math permutation problem and a full pipeline task requiring dataset generation, LoRA fine-tuning of a Gemma 2B model via Apple’s MLX framework, and serving a local web UI, stayed at perfect scores in both runs.
The cause of the discrepancy isn’t confirmed. It could reflect normal variance in one-shot generation, differences between the stealth serving setup and the production API, or a final release checkpoint that differs slightly from what was tested during the preview. Even at the lower score, 78.75% places GLM 5.3 Flash just below Claude Opus 4.8 and above Claude Opus 5, Kimi K3, DeepSeek V4 Pro, and its own predecessor GLM 5.2 on the same benchmark chart. The full-size GLM 5.3 model still scores highest overall, so Flash isn’t positioned to replace it, but as an MoE model with only 18 billion active parameters and a 15-cent API price, its benchmark position is notable.
Is GLM 5.3 Flash worth running locally?
The MIT license and MoE architecture make GLM 5.3 Flash one of the more practical large open models to self-host. Because only 18 billion of its 320 billion parameters activate per token, inference speed tracks closer to an 18B model even though memory requirements scale with the full parameter count. At 4-bit quantization the model requires roughly 180GB of memory. Community quantization projects were already working on 2-bit and 3-bit dynamic quants aimed at pushing that down toward the 100GB range, and MLX conversions for Apple Silicon appeared on Hugging Face shortly after release. ZAI’s own model card lists support across SGLang, vLLM, Transformers, KTransformers, and Unsloth for local or self-managed deployment.
The timing lines up with a new wave of high-memory local hardware. Apple’s newest Mac Studio configurations reach up to 512GB of unified memory with 1.2 terabytes per second of bandwidth, explicitly marketed for running models with hundreds of billions of parameters on-device. Xiaomi has also shown a compact AI box built around its own chips, with configurations reaching up to 160GB of unified memory and over a terabyte per second of near-memory bandwidth, demonstrated running a combined 120B-plus-3B model setup at 150 watts. Smaller unified-memory boxes in the DGX Spark class, around 128GB, round out this category. A model with a large total parameter count but a small active count is a good match for that hardware profile, since it needs the memory capacity these machines provide without demanding the raw compute a dense model of the same size would require.
Why did ZAI serve Ox Alpha on Chinese AI chips?
ZAI disclosed that the entire 44-trillion-token stealth preview ran on domestic Chinese AI chips rather than Nvidia GPUs, using a custom inference engine built on SGLang that the company says achieved roughly a 3x improvement in end-to-end serving performance. Running a frontier-adjacent model at that scale on non-Nvidia hardware, without users noticing any difference in quality or latency, suggests export controls on advanced chips aren’t preventing large-scale model serving the way they were once expected to. It also helps explain the economics behind the release: the efficiency gains from custom silicon and inference tooling likely made it feasible to give away that much free capacity during testing and to offer production API pricing as low as 15 cents per million input tokens afterward.
Frequently Asked Questions
What is GLM 5.3 Flash?
It’s an open-weight, natively multimodal model from ZAI (Zhipu AI), built as a mixture-of-experts model with 320 billion total parameters and 18 billion active parameters per token, released under an MIT license.
How is GLM 5.3 Flash related to Ox Alpha?
Ox Alpha was the anonymous name under which GLM 5.3 Flash was tested during a stealth preview period before ZAI’s official confirmation on August 26.
How much memory does GLM 5.3 Flash need to run locally?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
At 4-bit quantization it needs roughly 180GB of memory, with lower-bit dynamic quantizations from the community aiming to bring that down toward 100GB.
How does GLM 5.3 Flash compare to Claude Opus models?
On independent benchmark retesting it scored just below Claude Opus 4.8 and above Claude Opus 5, positioning it competitively despite its much lower API price and smaller active parameter count.
Is GLM 5.3 Flash better than the full GLM 5.3 model?
No. The full-size GLM 5.3 still scores highest on independent benchmarks. Flash trades some peak capability for dramatically lower cost, an MIT license, and practicality for local deployment.