The **BullSoar** Model GLM-5.3-Flash Officially Open Source: With 320B Parameters, Only 18B Activated, Priced at Just One-Tenth of GLM-5.2

Perceive Beating AI Newsletter: After Daytime Claiming of Ox Alpha, IntelliScript officially open-sourced GLM-5.3-Flash at night, with the model weights now available on Hugging Face under the MIT license. The total parameter size is 320 billion, but only 18 billion is activated per inference. It natively supports text, images, and video, making it the first native multimodal model in the GLM-5 series.

The officials stated that the overall performance of GLM-5.3-Flash exceeds that of GLM-5.2, with Programming and Agent evaluations approaching Claude Opus 4.8, but at only one-tenth of the price of GLM-5.2. It has also adopted a new base model, introducing sparse attention and a linear attention fusion architecture for the first time in the GLM main series, pretrained on 30T Tokens of multimodal data.

Prior to this, the community's DeepSWE small-sample test of Ox Alpha once achieved an 80% score, but that was only for 10 questions. When the sample size was increased, the score dropped to around 63%, and the testers actively corrected their initial claims. This score still places it in the top tier, but it is not as impressive as the initial 80%.

With the weights now open-sourced, developers can directly deploy it using frameworks such as vLLM, SGLang, KTransformers, etc., without the need to go through an anonymous model API.