Publication date: 27 August 2026
Z.ai has released GLM-5.3-Flash, an open-weight multimodal model with 320 billion total parameters and 18 billion active parameters per token. Z.ai says the model beats GLM-5.2 across its evaluations while operating at one tenth of the price.
Unlike GLM-5.3, which mainly improved post-training, GLM-5.3-Flash starts from a newly trained base model. Its architecture focuses on reducing long-context inference costs while adding native multimodal capabilities.
What can GLM-5.3-Flash do?
GLM-5.3-Flash is the first GLM-5-series model that Z.ai describes as natively multimodal. It combines sparse and linear attention, two approaches designed to reduce the amount of work required when processing long sequences.
The model has 320 billion parameters, with 18 billion active at a time. Z.ai says it supports context lengths up to one million tokens.
That makes the release particularly relevant for teams running long coding sessions, browser agents and computer-use workflows. Z.ai has already rolled it out to its Coding Plan and made the weights available for local deployment.
The Hugging Face release uses the permissive MIT licence and can be run through frameworks including vLLM and SGLang.
GLM-5.3-Flash benchmarks and cost
Z.ai claims GLM-5.3-Flash closes much of the gap with more expensive frontier models on coding and agentic workloads. These figures come from Z.ai’s own release evaluation and should be read as vendor-reported results.
| What is measured | GLM-5.3-Flash | Obvious alternative |
|---|---|---|
| Terminal Bench 2.1 | 84.3 | Claude Opus 4.8: 85.0 |
| DeepSWE 1.1 | 63.4 | Claude Opus 4.8: 58.0 |
| Toolathlon Verified | 78.4 | Claude Opus 4.8: 76.2 |
| AutomationBench | 48.8 | Claude Opus 4.8: 41.0 |
| OfficeQA Pro | 62.4 | Claude Opus 4.8: 48.9 |
Z.ai also reports an Artificial Analysis Intelligence Index score of 57 at a discounted cost of $0.045 per task. It describes overall pricing as one tenth of GLM-5.2, but the launch article does not provide a clean standard per-million-token price table.
That missing pricing detail matters. “One tenth the price” is useful positioning, but companies comparing API budgets need standard input, cached-input and output token rates.
How does GLM-5.3-Flash compare with GLM-5.2?
GLM-5.3-Flash is a more fundamental change than the name suggests. Z.ai says GLM-5.3 was based on the same base model as GLM-5.2, with improvements coming from post-training. GLM-5.3-Flash instead introduces a newly trained multimodal base and a different attention architecture.
On Z.ai’s DeepSWE evaluation, GLM-5.3-Flash scores substantially above GLM-5.2. Its AutomationBench advantage is larger still.
The caveat is familiar. Most of the detailed comparisons come from Z.ai itself. Independent evaluation is needed before assuming those differences will carry over to proprietary repositories, internal tools or browser workflows.
What this means
Worth testing now. Teams running high-volume coding agents should put GLM-5.3-Flash on a shortlist, particularly when token cost or self-hosting matters. The MIT licence also makes experimentation less complicated than models carrying custom commercial restrictions.
The first test should compare total task cost rather than token price alone. Measure how many attempts an agent needs, how long it stays on task and whether multimodal browser use improves completion rates. That will show whether the advertised efficiency translates into cheaper completed work.
For more information, visit the official announcement of GLM-5.3-Flash on the Z.ai website.
