This model is hosted in the United States and Europe. Like all other models on Ollama’s cloud, this model follows Ollama’s privacy policy with zero data retention.

GLM-5.3-Flash is the first natively multimodal model in Z.ai’s GLM-5 series. With 320B total parameters and just 18B active, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the cost, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

Before release, Z.ai tested GLM-5.3-Flash anonymously as ox-alpha where it became the most popular model of the week.

Key Features

• Agentic Coding: Built for sustained, multi-step coding in agent harnesses.
• Long Context: A 1M context window holds project-scale codebases and long-running sessions.
• Flexible Thinking: Reasoning is always on, with effort tunable per request across low, high, and max levels.
• Multimodal Input: Accepts image and video alongside text, for workflows that mix code with visual context.
• MIT Licensed

Competitive performance at Flash cost

Across six coding and agentic benchmarks, GLM-5.3-Flash consistently outperforms GLM-5.2, often by a wide margin: 63.4 vs. 46.2 on DeepSWE v1.1 and 48.8 vs. 26.2 on AutomationBench, while approaching Claude Opus 4.8 overall. On the Artificial Analysis Intelligence Index v4.1.1, it pushes the Pareto frontier at 57 points, a level of intelligence previously only available at roughly 10× the cost per task.

The same holds on Z.ai Code Bench v1.0 (run in Claude Code 2.1.207), where GLM-5.3-Flash beats GLM-5.2 at every effort level and, at Max effort, nearly matches Claude Opus 4.8 (29.0 vs. 29.5).

Built for efficient inference

GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. Compared with GLM-4.5, it has a similar total parameter count (320B vs. 355B) but nearly halves both the active parameters (18B vs. 32B) and the number of layers (45 vs. 92).

To keep attention cheap at long context, the model combines two kinds of attention. Linear attention captures local dependencies through state modeling, while sparse attention retrieves relevant global context through a lightweight indexer. IndexPool compresses four indexer key vectors into one, reducing the indexer’s latency and memory overhead at 1M-token context. Compared with GLM-5.3, GLM-5.3-Flash reduces attention compute by 3.0× and KV cache size by 4.4×. Together with a 30T-token multimodal pre-training corpus, these changes let the model deliver more intelligence with less compute.

Vision in the coding loop

For frontend, game development and 3D simulation, the final output is not code alone but an interface, an interaction or a world the user experiences, and many failures only show up once it is rendered, clicked or played. GLM-5.3-Flash is trained on visual coding trajectories in which the model interacts with an environment, inspects its own output and refines it iteratively, and, for frontend coding, with reinforcement learning from environment feedback. The same applies to computer use: the model decides when to look at the screen and uses what it sees to guide its next action.

Beyond coding, the model can read documents, spreadsheets, presentations, dashboards and interfaces directly as images rather than having them translated into text first, and check its own output against the intended visual result.

Benchmarks

Evaluation settings for each benchmark are listed in Z.ai’s blog post.

References

• Blog post: GLM-5.3-Flash: Frontier Intelligence, Flash Cost
• Hugging Face
• GLM-5 technical report