--- title: Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp... description: A demo comparing running the Qwen 3.8 Flash Next MoE model (125B parameters, ~6B activated, plus 51B engram parameters) locally on an Apple M5 Max machine... canonical: https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀 | daily.dev og:description: A demo comparing running the Qwen 3.8 Flash Next MoE model (125B parameters, ~6B activated, plus 51B engram parameters) locally on an Apple M5 Max machine... og:url: https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7 og:image: https://api.daily.dev/og/posts/WSYmcnMT7.png og:image:alt: Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀 og:image:width: 1200 og:image:height: 630 og:locale: en --- [Execute Automation](https://daily.dev/sources/executeautomation) [Watch video](https://api.daily.dev/r/WSYmcnMT7) # [Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀](https://api.daily.dev/r/WSYmcnMT7 "Go to post") A demo comparing running the Qwen 3.8 Flash Next MoE model (125B parameters, \~6B activated, plus 51B engram parameters) locally on an Apple M5 Max machine using two backends: llama.cpp with GGUF format versus MLX with the OMLX runtime. The MLX version, with lightning MTP (speculative decoding) enabled and SSD engram offloading, hits roughly 60 tokens per second and uses under 90GB of memory, compared to about 30 tokens per second and 120GB memory usage with llama.cpp/GGUF. The creator also runs a .NET 8 to .NET 10 migration task and an SVG generation task comparing the MoE Flash Next model against a dense Qwen 3.8 27B model, finding the MoE model faster on the coding task but noting the SVG test took longer yet produced a more detailed image. [#mixture-of-experts](/tags/mixture-of-experts "Check all #mixture-of-experts posts")[#llama-cpp](/tags/llama-cpp "Check all #llama-cpp posts") Aug 28•13m watch time Questions this post answers How much faster is MLX than llama.cpp for running the Qwen3.8 Flash Next model on Apple silicon? Running the Qwen3.8 Flash Next MoE model (125B parameters, 6B active, 51B engram parameters) via MLX with OMLX and lightning MTP enabled reaches roughly 60 tokens per second on an Apple M5 Max, compared to about 30 tokens per second using llama.cpp with the GGUF format on the same hardware, while also using less memory (under 90GB versus around 120GB). Anyone benchmarking local MoE model inference on Apple silicon can track these speed comparisons on daily.dev. What does the SSD engram offload feature do in OMLX 0.6.3? SSD engram offload keeps the large engram parameter table on disk instead of RAM and reads only the rows needed during inference, which reduces memory pressure but can slow prefill after context changes. In OMLX 0.6.3, this feature is combined with lightning MTP (a draft-token speculative decoding mechanism) to boost throughput for models like Qwen3.8 Flash Next. Developers tuning local inference memory tradeoffs can follow feature rollouts like this on daily.dev. Comment Bookmark Copy Share your thoughtsPost [](https://daily.dev/sources/executeautomation) [Execute Automation](https://daily.dev/sources/executeautomation "https://daily.dev/sources/executeautomation") 5 Followers • 20 Upvotes #### Would you recommend this post? Copy link WhatsApp Facebook X New Squad Copy linkShare with your friends ```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":"Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀","url":"https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7"},"datePublished":"2026-08-28T14:05:01.792Z","dateModified":"2026-08-29T17:06:19.581Z","description":"A demo comparing running the Qwen 3.8 Flash Next MoE model (125B parameters, ~6B activated, plus 51B engram parameters) locally on an Apple M5 Max machine...","image":"https://i.ytimg.com/vi/uCwvj3P3P9U/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/uCwvj3P3P9U/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Execute Automation","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":"Execute Automation","logo":"https://media.daily.dev/image/upload/s--Uh-EE9Cf--/f_auto,q_auto/v1780213688/logos/executeautomation?_a=BAMAMiWQ0","url":"https://daily.dev/sources/executeautomation"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"mixture-of-experts,llama-cpp","timeRequired":"PT13M","video":{"@type":"VideoObject","name":"Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀","description":"A demo comparing running the Qwen 3.8 Flash Next MoE model (125B parameters, ~6B activated, plus 51B engram parameters) locally on an Apple M5 Max machine...","thumbnailUrl":"https://i.ytimg.com/vi/uCwvj3P3P9U/sddefault.jpg","uploadDate":"2026-08-28T14:05:01.792Z","duration":"PT13M","url":"https://api.daily.dev/r/WSYmcnMT7","embedUrl":"https://www.youtube.com/embed/uCwvj3P3P9U"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Execute Automation","item":"https://daily.dev/sources/executeautomation"},{"@type":"ListItem","position":3,"name":"Qwen3.8-Flash-Next 125B: 60 TPS MLX vs 30 TPS llama.cpp on M5 Max 🚀"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/qwen3-8-flash-next-125b-60-tps-mlx-vs-30-tps-llama-cpp-on-m5-max--wsymcnmt7#faq","mainEntity":[{"@type":"Question","name":"How much faster is MLX than llama.cpp for running the Qwen3.8 Flash Next model on Apple silicon?","acceptedAnswer":{"@type":"Answer","text":"Running the Qwen3.8 Flash Next MoE model (125B parameters, 6B active, 51B engram parameters) via MLX with OMLX and lightning MTP enabled reaches roughly 60 tokens per second on an Apple M5 Max, compared to about 30 tokens per second using llama.cpp with the GGUF format on the same hardware, while also using less memory (under 90GB versus around 120GB). Anyone benchmarking local MoE model inference on Apple silicon can track these speed comparisons on daily.dev."}},{"@type":"Question","name":"What does the SSD engram offload feature do in OMLX 0.6.3?","acceptedAnswer":{"@type":"Answer","text":"SSD engram offload keeps the large engram parameter table on disk instead of RAM and reads only the rows needed during inference, which reduces memory pressure but can slow prefill after context changes. In OMLX 0.6.3, this feature is combined with lightning MTP (a draft-token speculative decoding mechanism) to boost throughput for models like Qwen3.8 Flash Next. Developers tuning local inference memory tradeoffs can follow feature rollouts like this on daily.dev."}}]} ```