Z.ai releases GLM-5.3-Flash, a 320B-parameter open-weight model under MIT license

Z.ai (Zhipu AI) has released GLM-5.3-Flash, a new open-weight large language model that pairs a large total parameter count with a small active-parameter footprint, aimed at cutting inference cost for coding and agentic workloads.

What's new

According to Z.ai's own developer documentation, "GLM-5.3-Flash has 320B total parameters with 18B activated." The model uses a mixture-of-experts design combined with a hybrid attention scheme, described in the docs as being among the "first open-source frontier model to combine sparse and linear attention," alongside a technique Z.ai calls "Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency."

Key specs from the documentation and the model's Hugging Face card:

• Architecture: 320B total parameters, 18B active per token (MoE), a newly trained base model rather than a fine-tune of a prior GLM checkpoint
• Context window: 1M tokens, with a maximum output of 128K tokens
• License: released under the MIT license, per the model card metadata on Hugging Face
• Multimodal support: native image, video, text, and file input, with visual capabilities "built into the coding loop" for agentic/coding use cases
• Availability: live now on the GLM Coding Plan, with Z.ai's documentation noting it ships with "3× the available quota compared with GLM-5.3" and reduced points consumption during off-peak usage

Z.ai had been quietly previewing the model under the codename "Ox Alpha" prior to the official release, before confirming it as the new GLM-5.3-Flash.

Context

GLM-5.3-Flash extends Z.ai's GLM line, which has positioned itself as a lower-cost, open-weight alternative to closed frontier models from Western labs. The company has increasingly targeted coding and agentic workloads specifically, building out products like ZCode and AutoClaw around its GLM models rather than only shipping raw weights. Releasing the model under the MIT license — a permissive license with minimal restrictions on commercial use and redistribution — continues Z.ai's pattern of open-sourcing frontier-scale models rather than gating them behind an API-only release, a strategy also used by DeepSeek, Alibaba's Qwen team, and Moonshot's Kimi models.

The MoE architecture, with only 18B of 320B total parameters active per token, is the same broad approach other open-weight labs have converged on this year to keep serving costs down while scaling total model capacity — trading a larger memory/storage footprint for cheaper, faster inference per request.

Why it matters

A 1M-token context window paired with a genuinely low active-parameter count is a combination aimed squarely at agentic and long-horizon coding tasks, where both long context retention and inference cost per step matter. Shipping it as open weights under MIT — rather than a more restrictive open license or API-only access — lowers the barrier for other companies to fine-tune, redistribute, or self-host the model without licensing negotiations.

The release also reinforces a broader trend: several of the most capable open-weight models now come from Chinese labs (Z.ai, DeepSeek, Alibaba, Tencent) rather than the US frontier labs, which have generally kept their most capable models closed. For developers optimizing for cost per token on agentic and coding workloads specifically, GLM-5.3-Flash adds another serious open option to evaluate alongside DeepSeek's and Qwen's recent releases.