Z.ai releases GLM-5.3-Flash with multimodal MIT weights

Z.ai released GLM-5.3-Flash on August 26 as the first natively multimodal model in the GLM-5 family. It accepts text, images, and video, exposes a one-million-token context window, and arrives through both an API and MIT-licensed weights.

The release is distinct from the larger, text-focused GLM-5.3. Z.ai previously tested the Flash model anonymously as ox-alpha through OpenCode and OpenRouter.

Key takeaways

• GLM-5.3-Flash is a 320-billion-parameter mixture-of-experts model with 18 billion active parameters.
• The model supports text, image, and video input within a one-million-token context window.
• Z.ai published FP8 and BF16 weights under the MIT license and provides the glm-5.3-flash API identifier.
• List pricing is $0.15 per million input tokens, $0.03 for cached input, and $0.50 for output.
• A 50% API promotion ends September 9, 2026, at 16:00 UTC; benchmark results remain primarily vendor-reported.

GLM-5.3-Flash combines vision, video, and agent work

The model uses a 320B-total, 18B-active sparse architecture with 45 layers and a combination of linear and sparse attention. Z.ai positions it for multimodal coding, browser or computer-use loops, document analysis, and agent workflows that must reason over visual evidence.

Unlike a text model with a separate vision adapter, Z.ai describes multimodal processing as native to the model’s training and architecture. The published model card supports images and video alongside text, while the API uses the same glm-5.3-flash identifier for hosted access.

Thinking mode cannot currently be disabled. That may improve difficult reasoning tasks, but it also affects latency and token consumption for workloads that would otherwise use a faster non-reasoning path.

Open weights and hosted access arrive together

Z.ai published official FP8 and BF16 repositories on Hugging Face under the MIT license. The model card documents deployment through SGLang, vLLM, TokenSpeed, and KTransformers, although the hardware requirements remain substantial even with only 18 billion parameters active per token.

Hosted access uses list prices of $0.15 per million input tokens, $0.03 per million cached tokens, and $0.50 per million output tokens. The temporary promotion cuts those rates in half until September 9 at 16:00 UTC. Developers should budget against the list price rather than treating the launch promotion as permanent.

The Flash model is also separate from the 753B GLM-5.3 text model. Existing GLM-5.3 integrations should not assume equivalent behavior, modalities, reasoning controls, or pricing merely because the names share a version number.

Benchmark gains need independent testing

Z.ai reports a DeepSWE score of 63.4, up from 46.2 for GLM-5.2, and an AutomationBench score of 48.8 versus 26.2. Its published multimodal results include 77.8 on MVBench and 78.0 on Chartography with tools.

Those numbers use vendor-selected harnesses and settings. LLM Stats has cataloged the results but explicitly labels them self-reported rather than independently verified. Its analysis also highlights weaker performance on BabyVision relative to Gemini 3.7 Flash and notes that comparisons across different agent harnesses are not automatically equivalent.

Teams evaluating the model should reproduce representative coding, document, and video tasks with their own tool permissions, context lengths, and cost measurements. The always-on reasoning mode makes latency and output-token usage especially important evaluation criteria.

What changes for model buyers

GLM-5.3-Flash gives developers a new option between smaller local vision models and expensive proprietary multimodal agents. Its combination of permissive weights, long context, native video input, and hosted pricing is the main release story; the launch benchmark table is supporting evidence, not a substitute for workload testing.

The next concrete milestones are independent benchmark results, practical serving reports for the 320B checkpoint, and the September 9 end of promotional pricing.

Source check

• Z.ai’s launch announcement describes the architecture, modalities, access, pricing, and vendor benchmark results.
• The official Hugging Face model card confirms the open-weight artifact, MIT license, and deployment guidance.
• LLM Stats’ independent launch analysis checks the specifications and pricing while separating vendor-reported benchmarks from independently verified results.