--- title: GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone... description: ZAI's GLM 5.3 and its smaller sibling Flash are open-weight models with an unusual sparse mixture-of-experts architecture: the flagship has 320B parameters but... canonical: https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone noticed | daily.dev og:description: ZAI's GLM 5.3 and its smaller sibling Flash are open-weight models with an unusual sparse mixture-of-experts architecture: the flagship has 320B parameters but... og:url: https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k og:image: https://api.daily.dev/og/posts/6SQf3Gz7k.png og:image:alt: GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone noticed 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. # GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone noticed **[Trends](https://daily.dev/sources/trends)** · 2 min read · 3 upvotes · 0 comments ## Summary ZAI's GLM 5.3 and its smaller sibling Flash are open-weight models with an unusual sparse mixture-of-experts architecture: the flagship has 320B parameters but activates only ~5% per token, combining linear and sparse attention with an 'index pool' context compression technique. Flash, built on a different base and RL-tuned for efficiency, costs roughly one-ninth the output token price and reportedly overtook DeepSeek in usage before most people realized what model it was. The flagship (744B total parameters) approaches frontier-level benchmarks with extended reasoning but requires thousands of dollars in hardware to self-host. The take: Flash is the practical, already-in-production product, while the flagship mainly proves the architecture scales. ## Content GLM 5.3 Flash showed up under a different name, nobody noticed, and then it quietly overtook DeepSeek in usage. That's the story. Zhipu AI's GLM 5.3 Flash is a 320B parameter sparse model that only activates around 5% of those parameters per token. The architecture is genuinely interesting: hybrid linear-plus-sparse attention handles nearby context cheaply, an "index pool" technique compresses stored context for faster long-session lookback, and the layer count was cut roughly in half from 92. The result is a model that runs fast enough that @nutlope watched it build a new page on his personal site in 37 seconds, unedited. The bigger GLM 5.3 sits at 744B parameters. Flash is the one getting all the attention. On pricing, Flash costs about 1/9th the output token price of the flagship. For agentic and coding work, that math is hard to argue with. @BHolmesDev put it plainly: "Open weight models have really caught up." He's planning a week of serious testing. @antirez ran his own benchmarks across cybersecurity, math, and QA, comparing Flash quants (Q2, Q4, mixed) against DeepSeek v4 Flash. The results are worth reading if you're deciding between the two for local inference. Unsloth already has GLM 5.3 Flash GGUFs in their desktop app with MTP enabled and faster long-context decoding out of the box. For vision work, @Sentdex is getting around 3Hz on tracking plus intelligence tasks with Flash alongside Qwen and DSV4F-vision. That's local, real-time, and actually usable. The caveats are real: you still need thousands of dollars of hardware to run the full model well, Flash uses more thinking tokens than competitors like Qwen, and it lags on some pure intelligence benchmarks. The flagship approaches frontier-level performance with extended reasoning time, but "approaches" is doing some work in that sentence. Still, the trajectory here is hard to ignore. A model that overtook DeepSeek in usage before most people knew its real name is not a model you dismiss. ## Questions this post answers ### What makes GLM 5.3 Flash cheaper to run than other coding-focused LLMs like DeepSeek? GLM 5.3 Flash is RL-tuned for efficiency rather than raw intelligence and costs about one-ninth the output token price of comparable models. It runs at over 200 tokens per second on Together's API, though it burns more thinking tokens than Qwen and lags on some intelligence benchmarks, making it attractive mainly for agentic and coding workloads where cost matters. _Track emerging low-cost coding models like GLM 5.3 Flash as pricing shifts on daily.dev._ ### How many parameters does the GLM 5.3 flagship model have and how many are active per token? GLM 5.3 has 320 billion total parameters but activates only around 5% of them per token, using a mix of linear and sparse attention plus an index-pool technique that compresses stored context for faster lookback in long sessions. Layer count was cut roughly in half from 92 compared to prior versions, enabling faster inference despite the large parameter count. _Developers comparing sparse-activation architectures can follow model internals like these on daily.dev._ --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-inference](https://daily.dev/tags/ai-inference), [#deepseek](https://daily.dev/tags/deepseek), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k) ```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":"GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone noticed","url":"https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k"},"datePublished":"2026-09-02T18:59:25.211Z","dateModified":"2026-09-07T18:14:40.904Z","description":"ZAI's GLM 5.3 and its smaller sibling Flash are open-weight models with an unusual sparse mixture-of-experts architecture: the flagship has 320B parameters but...","isAccessibleForFree":true,"articleSection":"Trends","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":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-inference,deepseek,mixture-of-experts","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"GLM 5.3 Flash quietly ate DeepSeek's lunch before anyone noticed"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/glm-5-3-flash-quietly-ate-deepseek-s-lunch-before-anyone-noticed-6sqf3gz7k#faq","mainEntity":[{"@type":"Question","name":"What makes GLM 5.3 Flash cheaper to run than other coding-focused LLMs like DeepSeek?","acceptedAnswer":{"@type":"Answer","text":"GLM 5.3 Flash is RL-tuned for efficiency rather than raw intelligence and costs about one-ninth the output token price of comparable models. It runs at over 200 tokens per second on Together's API, though it burns more thinking tokens than Qwen and lags on some intelligence benchmarks, making it attractive mainly for agentic and coding workloads where cost matters. Track emerging low-cost coding models like GLM 5.3 Flash as pricing shifts on daily.dev."}},{"@type":"Question","name":"How many parameters does the GLM 5.3 flagship model have and how many are active per token?","acceptedAnswer":{"@type":"Answer","text":"GLM 5.3 has 320 billion total parameters but activates only around 5% of them per token, using a mix of linear and sparse attention plus an index-pool technique that compresses stored context for faster lookback in long sessions. Layer count was cut roughly in half from 92 compared to prior versions, enabling faster inference despite the large parameter count. Developers comparing sparse-activation architectures can follow model internals like these on daily.dev."}}]} ```