Z.ai reveals Ox Alpha was GLM-5.3-Flash, an open multimodal model at one-tenth the price

Ox Alpha unmasked

For roughly a week, developers puzzled over a model listed as "Ox Alpha" on OpenCode and OpenRouter. It appeared on August 20 with no named owner, free to use, a one-million-token context window and native image and video input. According to dev.to, independent researchers fingerprinted its tokenizer and ran compression analyses, tracing it to Z.ai's GLM family with high confidence. On August 26 Z.ai confirmed the guess: Ox Alpha was GLM-5.3-Flash, published anonymously so the company could gather real-world feedback before the official launch.

A Mixture-of-Experts model shaped for cheap serving

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 line; dev.to notes the earlier GLM-5.2 and GLM-5.3 releases were text-only and drew user complaints for it. It is a Mixture-of-Experts design with 320 billion total parameters but only 18 billion active per token, and a 45-layer stack versus the 92 layers of the GLM-4.5 series. Z.ai compensated for the shallower shape with a 30-trillion-token multimodal pre-training corpus and a scaling technique it calls Manifold-Constrained Hyper-Connections.

The most notable engineering claim is the attention design. Z.ai describes this as the first open-source frontier model to combine sparse attention and linear attention in one architecture: linear attention captures local dependencies through state modeling, while a lightweight indexer retrieves global context. An IndexPool mechanism compresses four indexer key vectors into one via weighted pooling. Measured against GLM-5.3, Z.ai reports a 3x reduction in attention compute and a 4.4x smaller KV cache.

The benchmark picture, with caveats

Dev.to stresses that all the numbers are vendor-published, not independently verified, and that the comparison tables skip several strong models. With that caveat: Z.ai says GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at about $0.045 per task on the discounted tier. Artificial Analysis's own model page documents the index's composition, which spans nine evaluations including Terminal-Bench v2.1, GPQA Diamond and Humanity's Last Exam.

On coding and agentic work, Z.ai's figures show clear jumps over GLM-5.2: DeepSWE v1.1 rises from 46.2 to 63.4, ahead of Opus 4.8's 58.0; AutomationBench nearly doubles from 26.2 to 48.8; Terminal-Bench 2.1 reaches 84.3, just behind Opus 4.8's 85.0; Toolathlon Verified hits 78.4 and Agents' Last Exam 26.3, both close to the closed-model leaders. The clear weak spot is NL2Repo, where GLM-5.3-Flash scores 56.3 against Opus 4.8's 69.7. Generating an entire repository from a natural-language description, dev.to observes, remains firmly closed-frontier territory.

Vision built for the coding loop

Z.ai frames multimodality less as image processing and more as closing a feedback loop: a text-only model cannot see the broken layout its own frontend code produces. Training reportedly used self-visual judgment pipelines in which the model inspects its rendered output and refines it iteratively, plus reinforcement learning with environment feedback. The vision results split accordingly. Document and chart reasoning are strong — OfficeQA Pro at 62.4, well ahead of Opus 4.8's 48.9, and CharXiv Reasoning at 89.4, essentially tied with Opus 4.8. Raw video understanding lags: BabyVision sits at 53.4 versus Gemini 3.7 Flash's 70.9, and MVbench and MMVU both trail Gemini.

Running on domestic Chinese silicon

An easily missed infrastructure detail: throughout the preview week, the model was served on a cluster of domestically produced Chinese AI chips rather than NVIDIA hardware. Because those chips are constrained on memory capacity and bandwidth, especially at 1M-token contexts, Z.ai built a dedicated inference engine on top of SGLang with W8A8 quantization, hybrid INT8/FP8/BF16 cache quantization, Layer Split, and a disaggregated Encode-Prefill-Decode architecture. The company reports a 3x end-to-end serving improvement over its initial baseline, with per-token cost comparable to mainstream NVIDIA GPUs. Curiously, much of that optimization work was accelerated by a GLM-5.3-powered infrastructure agent that helped engineers write kernels and diagnose bottlenecks.

Pricing and access

The API is priced at $0.15 per million input tokens, $0.50 per million output tokens and $0.03 for cached input — roughly a tenth of GLM-5.3's $1.40/$4.40 rates. The GLM Coding Plan gains 3x the usable quota of GLM-5.3 at the same tier, and off-peak usage costs half the standard points. Weights are published on HuggingFace under the MIT license, with SGLang, vLLM and TokenSpeed supported for local deployment.

Why it matters

Three things lift this beyond a reveal stunt. The anonymous-testing playbook let Z.ai collect unbiased real-world feedback before launch, though the community's fingerprinting shows how short such anonymity's shelf life is. The economics matter more: an MIT-licensed, million-token-context, natively multimodal model at a tenth of its predecessor's price pushes frontier-adjacent capability within reach of ordinary projects. And the hardware story matters most of all — if a 1M-token model can be served at competitive cost on non-NVIDIA Chinese chips, the assumption that frontier inference requires NVIDIA GPUs weakens, and the hybrid attention design that enabled it is now open for anyone to study and reuse.