Zhipu releases GLM-5.3-Flash after anonymous Ox Alpha testing

The model behind Ox Alpha

Zhipu, operating internationally as Z.ai, identified the anonymously tested Ox Alpha model as GLM-5.3-Flash in August and released its weights. The official model card describes it as the first natively multimodal member of the GLM-5 series, with 320 billion total parameters and 18 billion active parameters.

Despite the shared 5.3 label, Flash starts from a newly trained base model. The card describes a hybrid of sparse and linear attention, manifold-constrained hyper-connections and a 30-trillion-token multimodal pretraining corpus. This is a change to the model's construction, not simply a faster serving setting for the larger GLM-5.3.

Vision inside a coding workflow

Z.ai's developer documentation lists image, video, text and file inputs, with text output, a one-million-token context and a maximum output of 128,000 tokens. It emphasizes tasks in which an agent inspects an interface or rendered result and then revises its code.

That feedback loop is different from producing code from a text specification alone. A screenshot can expose a misplaced element or a rendering failure that is not obvious in source code. The agent still needs a browser or other execution tools to obtain the image and apply changes; native vision supplies the model's ability to interpret that visual evidence.

Budget and interface choices affect the comparison

The hosted guide says thinking cannot be disabled for Flash, and recommends maximum reasoning effort when seeking its strongest results. That differs from assuming the Flash suffix means a short, non-reasoning response. The model's speed and final token bill depend on the actual inference settings and amount of work performed.

The public model card makes several evaluation budgets explicit. Its DeepSWE run uses mini-swe-agent with a six-hour timeout and 400,000-token context; Terminal-Bench 2.1 uses Claude Code with a six-hour timeout. Those are agent runs with external execution environments, not tests of a standalone reply. Comparing them with an interactive coding session requires preserving the harness and time allowance, not just the checkpoint name.

Z.ai's hosted API accepts visual inputs through image content blocks, and supports streamed tool calls and structured outputs. The Coding Plan instead has a points allowance: the guide lists three times the Flash quota relative to GLM-5.3, with reduced point consumption in off-peak periods. These subscription points are separate from a per-million-token API quote and do not make FlashX available through the plan.

Efficiency claims and access

The developer guide reports reductions in attention computation and KV-cache requirements relative to GLM-5.3. These are architecture-specific comparisons, distinct from the model card's advertised price comparison with GLM-5.2. They should not be combined into a single universal speedup.

The weights are available under MIT terms. The model card lists several local-serving frameworks and three reasoning-effort levels: low, high and max. It says benchmark reproduction uses max, which makes the chosen reasoning budget part of the performance comparison.

Z.ai also offers hosted access and includes Flash in its Coding Plan. Current documentation separately lists FlashX as a faster hosted variant and notes that its plan availability differs. Downloadable weights, an API endpoint and a subscription quota are therefore three different access arrangements, even when the underlying model names are closely related.