--- title: Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE?... description: Colibri's claim of running the 744B GLM 5.2 model on just 25 GB of RAM is technically true but practically misleading. The engine still requires ~370 GB of... canonical: https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE? Here’s Why | daily.dev og:description: Colibri's claim of running the 744B GLM 5.2 model on just 25 GB of RAM is technically true but practically misleading. The engine still requires ~370 GB of... og:url: https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck og:image: https://api.daily.dev/og/posts/Um00m1rcK.png og:image:alt: Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE? Here’s Why 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. # Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE? Here’s Why **[Medium](https://daily.dev/sources/medium_js)** · 4 min read · 1 upvotes · 0 comments ## Summary Colibri's claim of running the 744B GLM 5.2 model on just 25 GB of RAM is technically true but practically misleading. The engine still requires ~370 GB of NVMe storage, generates well under 1 token per second on most hardware, and only improves after repeated similar prompts warm up its cache. It lacks production-grade batching, making it unsuitable for multi-user serving. The complex setup — requiring knowledge of expert pinning, cache tuning, speculative decoding, and storage hierarchies — far exceeds tools like Ollama or LM Studio. While Colibri is an impressive engineering proof-of-concept for streaming MoE experts from disk, most developers seeking a fast local assistant will get better results from smaller 30B–70B models that fit entirely in RAM or VRAM. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much storage does Colibri need to run GLM 5.2 even though it only uses 25 GB of RAM? Around 370 GB of local NVMe storage is required just for the converted GLM 5.2 model, with even more temporary storage needed during the conversion process. The 25 GB RAM figure only covers memory usage, not the disk space needed to hold the streamed mixture-of-experts weights. _Anyone weighing storage versus RAM trade-offs for local model inference can track engineering critiques like this on daily.dev._ ### Why is Colibri's token generation speed so slow when running GLM 5.2? On a cold cache, generating each token can require reading nearly 11 GB of expert weights from disk, causing throughput to fall well below 1 token per second on many systems. Speed improves only with caching and speculative decoding, and reaching 1-2 tokens per second typically requires hundreds of gigabytes of RAM and multiple RTX 5090 GPUs. _Developers deciding whether disk-streaming inference is viable for their hardware can follow performance breakdowns like this on daily.dev._ ### Can Colibri serve multiple users simultaneously like vLLM or TensorRT-LLM? No, Colibri exposes an OpenAI-compatible API but processes only one generation at a time, unlike vLLM, TensorRT-LLM, or SGLang, which use continuous batching and parallel request processing to maximize GPU utilization. This makes Colibri unsuitable for production serving of multiple concurrent users. _Teams comparing inference engines for production serving can weigh trade-offs like these via daily.dev._ ## Similar posts on daily.dev - [I almost upgraded my GPU to run larger local LLMs, but this 8B model proved I didn't have to](https://daily.dev/posts/i-almost-upgraded-my-gpu-to-run-larger-local-llms-but-this-8b-model-proved-i-didn-t-have-to-gc24e505l) · XDA Developers · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck) ```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":"Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE? Here’s Why","url":"https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck"},"datePublished":"2026-07-21T18:48:34.555Z","dateModified":"2026-10-15T02:22:26.855Z","description":"Colibri's claim of running the 744B GLM 5.2 model on just 25 GB of RAM is technically true but practically misleading. The engine still requires ~370 GB of...","image":"https://media.daily.dev/image/upload/s--2-1xRawN--/f_auto/v1722860399/public/Placeholder%2011","thumbnailUrl":"https://media.daily.dev/image/upload/s--2-1xRawN--/f_auto/v1722860399/public/Placeholder%2011","isAccessibleForFree":true,"articleSection":"Medium","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":"Medium","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/medium","url":"https://daily.dev/sources/medium_js"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,mixture-of-experts","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Medium","item":"https://daily.dev/sources/medium_js"},{"@type":"ListItem","position":3,"name":"Colibri for Running GLM 5.2 in 25GBs RAM is a WASTE? Here’s Why"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/colibri-for-running-glm-5-2-in-25gbs-ram-is-a-waste-here-s-why-um00m1rck#faq","mainEntity":[{"@type":"Question","name":"How much storage does Colibri need to run GLM 5.2 even though it only uses 25 GB of RAM?","acceptedAnswer":{"@type":"Answer","text":"Around 370 GB of local NVMe storage is required just for the converted GLM 5.2 model, with even more temporary storage needed during the conversion process. The 25 GB RAM figure only covers memory usage, not the disk space needed to hold the streamed mixture-of-experts weights. Anyone weighing storage versus RAM trade-offs for local model inference can track engineering critiques like this on daily.dev."}},{"@type":"Question","name":"Why is Colibri's token generation speed so slow when running GLM 5.2?","acceptedAnswer":{"@type":"Answer","text":"On a cold cache, generating each token can require reading nearly 11 GB of expert weights from disk, causing throughput to fall well below 1 token per second on many systems. Speed improves only with caching and speculative decoding, and reaching 1-2 tokens per second typically requires hundreds of gigabytes of RAM and multiple RTX 5090 GPUs. Developers deciding whether disk-streaming inference is viable for their hardware can follow performance breakdowns like this on daily.dev."}},{"@type":"Question","name":"Can Colibri serve multiple users simultaneously like vLLM or TensorRT-LLM?","acceptedAnswer":{"@type":"Answer","text":"No, Colibri exposes an OpenAI-compatible API but processes only one generation at a time, unlike vLLM, TensorRT-LLM, or SGLang, which use continuous batching and parallel request processing to maximize GPU utilization. This makes Colibri unsuitable for production serving of multiple concurrent users. Teams comparing inference engines for production serving can weigh trade-offs like these via daily.dev."}}]} ```