--- title: MiniMax M3: Better than Claude and 20X Cheaper? (Open... description: MiniMax M3 is a new open-weights frontier model from Shanghai lab MiniMax, packing 428 billion total parameters (22-23B active via mixture-of-experts) and up... canonical: https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: MiniMax M3: Better than Claude and 20X Cheaper? (Open Source) | daily.dev og:description: MiniMax M3 is a new open-weights frontier model from Shanghai lab MiniMax, packing 428 billion total parameters (22-23B active via mixture-of-experts) and up... og:url: https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9 og:image: https://api.daily.dev/og/posts/eqM63NsK9.png og:image:alt: MiniMax M3: Better than Claude and 20X Cheaper? (Open Source) 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. # MiniMax M3: Better than Claude and 20X Cheaper? (Open Source) **[The Coding Gopher](https://daily.dev/sources/codinggopher)** · 9 min read · 42 upvotes · 7 comments ## Summary MiniMax M3 is a new open-weights frontier model from Shanghai lab MiniMax, packing 428 billion total parameters (22-23B active via mixture-of-experts) and up to 1 million tokens of context. It uses a custom sparse attention scheme (MSA) that groups tokens into 128-token hardware-aligned blocks, cutting per-token compute cost at 1M context to roughly 1/20 of the prior M2 model, with 9.7x faster prefill and 15.6x faster decoding. It supports native multimodal understanding including video up to 30 minutes. On benchmarks it scores 59% on SWE-Bench Pro, 66% on Terminal-Bench 2.1, and 83.5% on BrowseComp, beating Claude Opus 4.7's 79.3%. API pricing on OpenRouter launched at 30 cents per million input tokens. Running it locally requires roughly 855GB unquantized, or a minimum of 133GB unified memory even after aggressive 1-bit quantization, putting local hosting out of reach for most individual developers. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How many parameters does MiniMax M3 have and how many are active per token? MiniMax M3 has roughly 428 billion total parameters but uses a mixture-of-experts architecture that activates only about 22 to 23 billion parameters per token. This sparse activation preserves the reasoning quality associated with a much larger dense model while keeping inference throughput high compared to running the full parameter count on every token. _Engineers sizing hardware for MoE model deployment can track architecture breakdowns like this one on daily.dev._ ### How much memory do you need to run MiniMax M3 locally? Running MiniMax M3 locally requires a minimum of 133 GB of unified memory (RAM plus VRAM) even after shrinking the unquantized 855 GB BF16 weights down to about 128 GB using extreme 1-bit GGUF quantization. For decent performance with 3-bit or 4-bit quantization, 164 to 270 GB of total memory is needed, putting local hosting out of reach for most consumer hardware. _Developers weighing local hosting against cloud APIs can follow hardware requirement breakdowns like this on daily.dev._ ### How does MiniMax Sparse Attention (MSA) reduce inference cost for long context windows? MiniMax Sparse Attention replaces full attention with an index branch that scores memory and pulls only the top-K most relevant tokens into active SRAM, grouped into hardware-aligned 128-token blocks read sequentially rather than scattered across memory. This cuts per-token compute cost at 1 million tokens to about 1/20 of the previous M2 model, delivering a 9.7x prefill speedup and 15.6x decoding speedup. _Developers optimizing long-context inference costs can track architecture explainers like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@jepudev2** · 11 upvotes > tl;dr: its not **@dash15tonon** · 4 upvotes > You have the official benchmark here [MiniMax M3 - Coding & Agentic Frontier, 1M Context, Multimodal | MiniMax](https://www.minimax.io/models/text/m3) > > I don't think is better than Claude but they can improve it and in a near future they can certainly beat anthropic **@hardik15** · 4 upvotes > What's the use of this comparison and publishing if it's out of reach for indie dev and most of small companies as well. Even after investing this much on hardware the speed and output is still what they were expecting. > > > Also the future versions will need more and more H/w power/configuration unless they achieve SLM in very effective and accurate way **@jeremysylvis** · 3 upvotes > I'd like to see it compared to Sonnet 5 - that seems to be what a lot of orgs with Claude EAs are targeting for general enterprise use and general software engineering. > > > M3 _seems_ to be a strong contender in everything except the long horizon stuff, and it seems to be doing as well as anyone that isn't Anthropic/OpenAI. It's certainly got an interesting price point. **@jayelsheikh** · 0 upvotes > Its been no 1 used model on thinkreview --- Tags: [#open-source](https://daily.dev/tags/open-source), [#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/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9) ```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":"MiniMax M3: Better than Claude and 20X Cheaper? (Open Source)","url":"https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9"},"datePublished":"2026-08-25T16:05:30.633Z","dateModified":"2026-09-14T06:08:02.774Z","description":"MiniMax M3 is a new open-weights frontier model from Shanghai lab MiniMax, packing 428 billion total parameters (22-23B active via mixture-of-experts) and up...","image":"https://i.ytimg.com/vi/d0P7xoO2gD4/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/d0P7xoO2gD4/sddefault.jpg","isAccessibleForFree":true,"articleSection":"The Coding Gopher","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":"The Coding Gopher","logo":"https://media.daily.dev/image/upload/s--pv2IKQMG--/f_auto/v1736753456/logos/codinggopher","url":"https://daily.dev/sources/codinggopher"},"commentCount":7,"discussionUrl":"https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":42},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":7}],"keywords":"open-source,llm,mixture-of-experts","timeRequired":"PT9M","video":{"@type":"VideoObject","name":"MiniMax M3: Better than Claude and 20X Cheaper? (Open Source)","description":"MiniMax M3 is a new open-weights frontier model from Shanghai lab MiniMax, packing 428 billion total parameters (22-23B active via mixture-of-experts) and up...","thumbnailUrl":"https://i.ytimg.com/vi/d0P7xoO2gD4/sddefault.jpg","uploadDate":"2026-08-25T16:05:30.633Z","duration":"PT9M","url":"https://api.daily.dev/r/eqM63NsK9","embedUrl":"https://www.youtube.com/embed/d0P7xoO2gD4"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The Coding Gopher","item":"https://daily.dev/sources/codinggopher"},{"@type":"ListItem","position":3,"name":"MiniMax M3: Better than Claude and 20X Cheaper? (Open Source)"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9","comment":[{"@type":"Comment","text":"tl;dr: its not","datePublished":"2026-08-25T23:17:38.412Z","url":"https://daily.dev/posts/eqM63NsK9#c-hnwEbBMEV","author":{"@type":"Person","name":"jepudev","url":"https://daily.dev/jepudev2","image":"https://avatars.githubusercontent.com/u/71714253?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":11}},{"@type":"Comment","text":"You have the official benchmark here MiniMax M3 - Coding & Agentic Frontier, 1M Context, Multimodal | MiniMax\nI don’t think is better than Claude but they can improve it and in a near future they can certainly beat anthropic","datePublished":"2026-08-26T07:46:48.409Z","dateModified":"2026-08-26T07:48:51.677Z","url":"https://daily.dev/posts/eqM63NsK9#c-OK9iZthsn","author":{"@type":"Person","name":"David15tonon","url":"https://daily.dev/dash15tonon","image":"https://media.daily.dev/image/upload/s--ODg7gW3B--/f_auto/v1779805078/avatars/avatar_I7Oy0mpyXs8dOSJPf3bPC?_a=BAMAMiWQ0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4}},{"@type":"Comment","text":"What’s the use of this comparison and publishing if it’s out of reach for indie dev and most of small companies as well. Even after investing this much on hardware the speed and output is still what they were expecting.\nAlso the future versions will need more and more H/w power/configuration unless they achieve SLM in very effective and accurate way","datePublished":"2026-08-26T06:10:10.474Z","url":"https://daily.dev/posts/eqM63NsK9#c-dpib7Hzab","author":{"@type":"Person","name":"Hardik","url":"https://daily.dev/hardik15","image":"https://media.daily.dev/image/upload/s--O0TOmw4y--/f_auto/v1715772965/public/noProfile"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4}},{"@type":"Comment","text":"I’d like to see it compared to Sonnet 5 - that seems to be what a lot of orgs with Claude EAs are targeting for general enterprise use and general software engineering.\nM3 seems to be a strong contender in everything except the long horizon stuff, and it seems to be doing as well as anyone that isn’t Anthropic/OpenAI. It’s certainly got an interesting price point.","datePublished":"2026-08-26T13:30:04.157Z","url":"https://daily.dev/posts/eqM63NsK9#c-5F6lQD5wh","author":{"@type":"Person","name":"Jeremy Sylvis","url":"https://daily.dev/jeremysylvis","image":"https://avatars.githubusercontent.com/u/58440072?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}},{"@type":"Comment","text":"Its been no 1 used model on thinkreview","datePublished":"2026-08-29T02:18:49.607Z","url":"https://daily.dev/posts/eqM63NsK9#c-QD2VWCpTj","author":{"@type":"Person","name":"Jay Elsheikh","url":"https://daily.dev/jayelsheikh","image":"https://lh3.googleusercontent.com/a/ACg8ocKOYmya6W94Us-PkLZOWFTNLlErDLzZOiX9BAA-7bRQCMAAezCOkg=s96-c"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/minimax-m3-better-than-claude-and-20x-cheaper-open-source--eqm63nsk9#faq","mainEntity":[{"@type":"Question","name":"How many parameters does MiniMax M3 have and how many are active per token?","acceptedAnswer":{"@type":"Answer","text":"MiniMax M3 has roughly 428 billion total parameters but uses a mixture-of-experts architecture that activates only about 22 to 23 billion parameters per token. This sparse activation preserves the reasoning quality associated with a much larger dense model while keeping inference throughput high compared to running the full parameter count on every token. Engineers sizing hardware for MoE model deployment can track architecture breakdowns like this one on daily.dev."}},{"@type":"Question","name":"How much memory do you need to run MiniMax M3 locally?","acceptedAnswer":{"@type":"Answer","text":"Running MiniMax M3 locally requires a minimum of 133 GB of unified memory (RAM plus VRAM) even after shrinking the unquantized 855 GB BF16 weights down to about 128 GB using extreme 1-bit GGUF quantization. For decent performance with 3-bit or 4-bit quantization, 164 to 270 GB of total memory is needed, putting local hosting out of reach for most consumer hardware. Developers weighing local hosting against cloud APIs can follow hardware requirement breakdowns like this on daily.dev."}},{"@type":"Question","name":"How does MiniMax Sparse Attention (MSA) reduce inference cost for long context windows?","acceptedAnswer":{"@type":"Answer","text":"MiniMax Sparse Attention replaces full attention with an index branch that scores memory and pulls only the top-K most relevant tokens into active SRAM, grouped into hardware-aligned 128-token blocks read sequentially rather than scattered across memory. This cuts per-token compute cost at 1 million tokens to about 1/20 of the previous M2 model, delivering a 9.7x prefill speedup and 15.6x decoding speedup. Developers optimizing long-context inference costs can track architecture explainers like this on daily.dev."}}]} ```