Technical Specifications and Architecture

Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series and the first open-weight release utilizing the glm5_next architecture. The model is a causal language model with a vision encoder, featuring 320 billion total parameters with 18 billion activated.

The architecture employs a hybrid approach, utilizing 45 layers composed of repeating blocks of three KDA linear attention layers followed by one DeepSeek-style sparse attention layer. The sparse layers utilize a lightning indexer with a top-k budget of 2048 tokens, a design intended to reduce the cost of serving long-context windows. Additionally, the model incorporates Manifold-Constrained Hyper-Connections (mHC) to widen residual streams and improve scaling efficiency.

The vision encoder consists of 24 layers with a 448x448 image size and a 14-pixel patch size. According to technical documentation from Z.ai, the model was trained on a 30-trillion-token multimodal corpus, allowing for both image and video tokens to be included in the vocabulary. The model supports a context window of up to 1,048,576 tokens.

Performance Claims and Benchmarks

Z.ai claims that GLM-5.3-Flash outperforms its predecessor, GLM-5.2, at approximately one-tenth of the cost. The company asserts that the model approaches the performance of Claude Opus 4.8 on coding and agentic benchmarks.

Unverified user reports on Reddit suggest the model is highly capable in agentic coding tasks, with some users noting significant “thinking” time during complex reasoning processes. However, some community members have expressed skepticism regarding these performance claims, suggesting that benchmarks may be “maxed” and that real-world performance may vary.

The model is being released in two primary formats: an FP8 version (e4m3 with dynamic activation scaling) and a BF16 version.

Market Context and “Ox Alpha” Pre-release

Prior to the official release, the model was tested anonymously under the name “Ox Alpha” on platforms such as OpenRouter and OpenCode. According to Z.ai, the Ox Alpha version became the most popular model of the week during its testing phase. Notably, the company stated that all traffic for Ox Alpha was served on Chinese AI chips, a detail that has drawn attention regarding the development of independent compute capabilities in China.

While some community speculation on Reddit attempted to link “Ox Alpha” to Google’s Gemini model, these claims remain unverified and are largely viewed by industry observers as speculative.

Deployment and Availability

The model weights are available via Hugging Face and ModelScope. For inference, official support has been verified for vLLM (version 0.27.0+) and SGLang, with the latter offering configurations for various hardware including H100, H200, and B200 series GPUs. Other supported frameworks include TokenSpeed and KTransformers.

The model is released under the MIT license.