This AI Has 320 Billion Parameters. It Barely Uses Them.

Name: This AI Has 320 Billion Parameters. It Barely Uses Them.
Uploaded: 2026-09-01T09:04:56.000Z
Duration: 5 min 11 s
Channel: Two Minute Papers
Description: GLM 5.3 Flash, an AI model with 320 billion parameters that efficiently utilizes its capabilities, and highlights implications for AI performance and efficiency. The channel, Two Minute Papers, regularly discusses AI topics, making it relevant to the AI landscape.

What you'll learn

GLM 5.3 Flash has 320 billion parameters, of which about 20 billion are activated per token.
The model was released anonymously as Ox Alpha and quickly overtook even DeepSeek in usage.
Two techniques, linear attention and an index pool, lower compute costs and memory use.
According to Zsolnai-Fehér, the top model Fable could be surpassed by free AI within this year.
The number of layers was roughly halved, from 92 in the previous series to about half.

Frequently asked questions

What makes GLM 5.3 Flash special?

The Zhipu AI model has 320 billion parameters but activates only about 20 billion per token. That allows it to run on your own computer without a subscription, according to the presenter.

Which techniques make the model efficient?

Zhipu uses linear attention, which summarizes nearby context into a small package, and an index pool that compresses stored context before searching. That lets the model look further back with less memory and compute.

Can GLM 5.3 Flash run at home?

Yes, but not on an ordinary machine. Zsolnai-Fehér mentions hardware in the order of thousands of euros, and even that is not always enough for the big GLM 5.3 model. Compressed versions run on more devices but can stall into repeating output.

When was GLM 5.3 Flash released?

The model appeared anonymously as Ox Alpha around August 20, 2026, and Bloomberg confirmed on August 26, 2026 that it was GLM 5.3 Flash by Zhipu. The flagship GLM 5.3 came earlier, on August 14, 2026.

Topics

glm-5.3-flash large-language-models ai-efficiency model-architecture

In this video

Has

Sources

What is known about this topic outside the broadcast, and where it says so.

Description from the channel

❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers 📝 GLM 5.3 Flash: https://z.ai/blog/glm-5.3-flash Sources: https://x.com/louszbd/status/2092694163104113016 https://x.com/semianalysis_/status/2092623833630998556 https://x.com/skalskip92/status/2092748209802154201 https://x.com/louszbd/status/2093047548550525165 https://x.com/atomic_chat_hq/status/2093433913238552712 https://x.com/holytrinity/status/2094093933584257334 https://x.com/KinasRemek/status/2090081611832295581/video/1 https://x.com/AiXsatoshi/status/2093679264013181119/video/1 https://x.com/AiXsatoshi/status/2093353322921263389/video/1 https://x.com/stevibe/status/2092655031040565252 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi