<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/qIONyfhmr" -->

---
title: This AI Has 320 Billion Parameters. It Barely Uses Them.
description: GLM 5.3 and its smaller sibling GLM 5.3 Flash are new open-weight AI models that pack 320 billion parameters but activate only about 5% per token, using linear...
canonical: https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: This AI Has 320 Billion Parameters. It Barely Uses Them. | daily.dev
og:description: GLM 5.3 and its smaller sibling GLM 5.3 Flash are new open-weight AI models that pack 320 billion parameters but activate only about 5% per token, using linear...
og:url: https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr
og:image: https://api.daily.dev/og/posts/qIONyfhmr.png
og:image:alt: This AI Has 320 Billion Parameters. It Barely Uses Them.
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

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

**[Two Minute Papers](https://daily.dev/sources/twoninutepapers)** · 5 min read · 3 upvotes · 0 comments

## Summary

GLM 5.3 and its smaller sibling GLM 5.3 Flash are new open-weight AI models that pack 320 billion parameters but activate only about 5% per token, using linear attention plus sparse attention to summarize nearby context cheaply, and a technique called 'index pool' to compress stored context for faster long-session lookback. The layer count was roughly halved from 92. Flash quickly overtook DeepSeek in usage after being stealthily released under another name, and both models reportedly approach frontier-level performance on some benchmarks when given extended reasoning time, though still requiring thousands of dollars of hardware to run well.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://www.youtube.com/watch?v=w9RDunJACkc>

## Questions this post answers

### How many of GLM 5.3's 320 billion parameters are actually activated per token?

Only about 5% of the model's 320 billion parameters are activated per token, with roughly 95% left inactive for any given inference pass. This sparse activation, combined with roughly halving the number of layers from a previous 92, lets the model run far cheaper than its full parameter count would suggest.

_Developers weighing GPU budgets against model size can track efficiency-focused releases like GLM 5.3 on daily.dev._

### What is linear attention and how does GLM 5.3 use it?

Linear attention is a technique GLM 5.3 uses to summarize nearby context into a compact package instead of comparing every token to every other token, making it dramatically cheaper than even sparse attention, which the model also employs. Together these methods let the model handle long conversations, large codebases, or lengthy documents without the usual runaway compute and memory costs.

_Anyone comparing attention mechanisms for long-context workloads can follow model architecture breakdowns on daily.dev._

### What is the index pool feature in GLM 5.3 used for?

Index pool is GLM 5.3's method of compressing an index of stored context before searching through it, so the model can look back further in long documents, codebases, or discussions while using less memory and compute. It addresses the common problem where an AI's responses degrade as a conversation or context window grows longer.

_Teams building long-context AI applications can keep up with context-handling techniques like this on daily.dev._

---

Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#ai-inference](https://daily.dev/tags/ai-inference)

[View this post on daily.dev](https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"This AI Has 320 Billion Parameters. It Barely Uses Them.","url":"https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr"},"datePublished":"2026-09-01T09:06:51.007Z","dateModified":"2026-09-14T06:45:54.876Z","description":"GLM 5.3 and its smaller sibling GLM 5.3 Flash are new open-weight AI models that pack 320 billion parameters but activate only about 5% per token, using linear...","image":"https://i.ytimg.com/vi/w9RDunJACkc/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/w9RDunJACkc/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Two Minute Papers","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Two Minute Papers","logo":"https://media.daily.dev/image/upload/s--8KXsSJ5q--/f_auto/v1711188923/logos/twoninutepapers","url":"https://daily.dev/sources/twoninutepapers"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,ai-inference","timeRequired":"PT5M","video":{"@type":"VideoObject","name":"This AI Has 320 Billion Parameters. It Barely Uses Them.","description":"GLM 5.3 and its smaller sibling GLM 5.3 Flash are new open-weight AI models that pack 320 billion parameters but activate only about 5% per token, using linear...","thumbnailUrl":"https://i.ytimg.com/vi/w9RDunJACkc/sddefault.jpg","uploadDate":"2026-09-01T09:06:51.007Z","duration":"PT5M","url":"https://api.daily.dev/r/qIONyfhmr","embedUrl":"https://www.youtube.com/embed/w9RDunJACkc"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Two Minute Papers","item":"https://daily.dev/sources/twoninutepapers"},{"@type":"ListItem","position":3,"name":"This AI Has 320 Billion Parameters. It Barely Uses Them."}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/this-ai-has-320-billion-parameters-it-barely-uses-them--qionyfhmr#faq","mainEntity":[{"@type":"Question","name":"How many of GLM 5.3's 320 billion parameters are actually activated per token?","acceptedAnswer":{"@type":"Answer","text":"Only about 5% of the model's 320 billion parameters are activated per token, with roughly 95% left inactive for any given inference pass. This sparse activation, combined with roughly halving the number of layers from a previous 92, lets the model run far cheaper than its full parameter count would suggest. Developers weighing GPU budgets against model size can track efficiency-focused releases like GLM 5.3 on daily.dev."}},{"@type":"Question","name":"What is linear attention and how does GLM 5.3 use it?","acceptedAnswer":{"@type":"Answer","text":"Linear attention is a technique GLM 5.3 uses to summarize nearby context into a compact package instead of comparing every token to every other token, making it dramatically cheaper than even sparse attention, which the model also employs. Together these methods let the model handle long conversations, large codebases, or lengthy documents without the usual runaway compute and memory costs. Anyone comparing attention mechanisms for long-context workloads can follow model architecture breakdowns on daily.dev."}},{"@type":"Question","name":"What is the index pool feature in GLM 5.3 used for?","acceptedAnswer":{"@type":"Answer","text":"Index pool is GLM 5.3's method of compressing an index of stored context before searching through it, so the model can look back further in long documents, codebases, or discussions while using less memory and compute. It addresses the common problem where an AI's responses degrade as a conversation or context window grows longer. Teams building long-context AI applications can keep up with context-handling techniques like this on daily.dev."}}]}
```

