[HN Gopher] MiniMax-M1 open-weight, large-scale hybrid-attention...
       ___________________________________________________________________
        
       MiniMax-M1 open-weight, large-scale hybrid-attention reasoning
       model
        
       Author : danboarder
       Score  : 315 points
       Date   : 2025-06-18 06:53 UTC (16 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | swyx wrote:
       | 1. this is apparently MiniMax's "launch week" - they did M1 on
       | Monday and Hailuo 2 on Tuesday
       | (https://news.smol.ai/issues/25-06-16-chinese-models). remains to
       | be seen if they can keep up the pace of model releases for the
       | rest of this week - these 2 were big ones, they aren't yet known
       | for much else beyond llm and video models. just watch
       | https://x.com/MiniMax__AI for announcements.
       | 
       | 2. minimax m1's tech report is worthwhile:
       | https://github.com/MiniMax-AI/MiniMax-M1/blob/main/MiniMax_M...
       | while they may not be the SOTA open weights model, they do make
       | some very big/notable claims on lightning attention and their
       | GRPO variant (CISPO).
       | 
       | (im unaffiliated, just sharing what ive learned so far since no
       | comments have been made here yet
        
         | behnamoh wrote:
         | > they did M1 on Monday and Hailuo 2 on Tuesday
         | 
         | It would've been fun to see them name their models like Apple
         | chips: M1, M1 Pro, M1 Ultra.
        
           | plufz wrote:
           | Yeah MiniMax M1 certainly directed my thoughts to Mac mini
           | M1. :)
        
       | noelwelsh wrote:
       | A few thoughts:
       | 
       | * A Singapore based company, according to LinkedIn. There doesn't
       | seem to be much of a barrier to entry to building a very good
       | LLM.
       | 
       | * Open weight models + the development of Strix Halo / Ryzen AI
       | Max makes me optimistic that running great LLMs locally will be
       | relatively cheap in a few years.
        
         | rfoo wrote:
         | > A Singapore based company, according to LinkedIn
         | 
         | Nah, this is a Shanghai-based company.
        
         | manc_lad wrote:
         | It seems more and more like an inevitability we will run models
         | locally. Exciting and concerning implications.
         | 
         | If anyone has any suggestions of people thinking about this
         | space they respect, I'd love to listen to more ideas and
         | thoughts on the developments.
        
           | noelwelsh wrote:
           | I think the main limitation, right now, is hardware. For GPUs
           | the main limit is the VRAM available on consumer models. CPUs
           | have plenty of memory but don't have the bandwidth or vector
           | compute power for LLMs. This is why I think the Strix Halo is
           | so exciting: it has bandwidth + compute power plus a lot of
           | memory. It's not quite where it needs to be to replace a
           | dedicated GPU, but in a few iterations it could be.
           | 
           | I'm interested in other opinions. I'm no expert on this
           | stuff.
        
             | jb1991 wrote:
             | How does the shared memory model for GPUs on Apple Silicon
             | factor into this? These are technically consumer grade and
             | not very expensive, but they can offer a huge amount of
             | memory since all the memory is shared between CPU and GPU,
             | even a midtier machine can easily have 100 GB of GPU
             | memory.
        
               | noelwelsh wrote:
               | If you squint the M4 is the same as the Strix Halo. The
               | M4 has roughly
               | 
               | * double the bandwidth;
               | 
               | * half the compute; and
               | 
               | * double the price for comparable memory (128GB)
               | 
               | compared to the Strix Halo.
               | 
               | I'm more interested in the AMD chips because of cost
               | plus, while I have an Apple laptop, I do most of my work
               | on a Linux desktop. So a killer AMD chip works better for
               | me. If you don't mind paying the Apple tax then a Mac is
               | a viable option. I'm not sure on the software side of
               | LLMs on Apple Silicon but I cannot imagine it's unusable.
               | 
               | An example of desktop with the Strix Halo is the
               | Framework desktop (AI Max+ 395 is the marketing name for
               | the Strix Halo chip with the most juice):
               | https://frame.work/gb/en/products/desktop-diy-amd-
               | aimax300/c...
        
               | justincormack wrote:
               | Apple has machines with 2x and about 3x the Strix Halo
               | bandwidth by doubling up the memory buses. These get
               | expensive though.
        
               | ezschemi wrote:
               | I am also very interested in AMD's Strix Halo for running
               | LLMs locally. For that I have a Framework Desktop in
               | order (batch 1!). Alex Ziskind on Youtube does videos
               | comparing Strix Halo, M4 Mac mini and MacBook Pro, Nvidia
               | 5090, etc. including power consumption. The only downside
               | is one has to pull out the numbers from the videos,
               | there's no tables or anything. Here is the recent video
               | with testing Strix Halo and a Mac mini:
               | https://www.youtube.com/watch?v=B7GDr-VFuEo
        
           | pantulis wrote:
           | Honest question: what is the concerning aspect to it?
        
           | psychoslave wrote:
           | I don't know, what's worst with people running LLM locally
           | compared to running any software locally?
           | 
           | There is nothing fundamentally new in having freedom in edge
           | of societies. Yes it can lead to horrible situation, like
           | someone kill neighbors, using the single handable bright new
           | tool available to all. But that's far less of a concern than
           | having the powerful new tool staying in full concentrated
           | control of the greediest humans out there, who will gladly
           | escalate any hindrance to genocide whenever something doesn't
           | fit their perspective.
        
         | ulfw wrote:
         | They are a Chinese company based out of the city of Shanghai,
         | not Singapore.
         | 
         | They're also planning to IPO at HKEX in Hong Kong soon
         | 
         | https://www.scmp.com/tech/tech-trends/article/3314819/deepse...
        
           | 7moritz7 wrote:
           | I'll keep an eye out on that ipo
        
         | freeqaz wrote:
         | In a linked Twitter post[0], they trained this for $500k-ish. I
         | wonder how?
         | 
         | > RL at unmatched efficiency: trained with just $534,700
         | 
         | 0: https://x.com/MiniMax__AI/status/1934637031193514237
        
       | vintermann wrote:
       | "We publicly release MiniMax-M1 at this https url" in the arxiv
       | paper, and it isn't a link to an empty repo!
       | 
       | I like these people already.
        
       | npteljes wrote:
       | This is stated nowhere on the official pages, but it's a Chinese
       | company.
       | 
       | https://en.wikipedia.org/wiki/MiniMax_(company)
        
         | iLoveOncall wrote:
         | Why would you expect them to mention that on their project's
         | page?
        
           | noelwelsh wrote:
           | 1. It's conventional to do so.
           | 
           | 2. It's a legal requirement in some jurisdictions (e.g.
           | https://www.gov.uk/running-a-limited-company/signs-
           | stationer...)
           | 
           | 3. It's useful for people who may be interested in applying
           | for jobs
        
             | iLoveOncall wrote:
             | 1. No it's not. Top GitHub repository from Google as an
             | example: https://github.com/google/material-design-icons I
             | think you'd actually be hard pressed to find a single
             | repository where the company that owns it lists where they
             | are registered.
             | 
             | 2. This is a requirement for companies registered in the
             | UK. You should also read your own link, it doesn't say
             | anything about the company's presence on 3rd party
             | websites.
             | 
             | 3. This is such a remote reason it's laughable, there are
             | plenty more things that are more relevant to potential job
             | applications, such as whether they are hiring at all or
             | not.
             | 
             | You just want them to mention it because it's a Chinese
             | company. If they were American, Mexican, German or
             | Zimbabwean you wouldn't give the slightest fuck.
        
               | noelwelsh wrote:
               | OP said "official pages", which I took to mean the
               | company website: https://www.minimax.io/
               | 
               | Also, thanks for putting words in my mouth. If they were
               | Mexican or Zimbabwean I would find it very interesting to
               | see a roughly SOtA model coming from that country.
        
               | npteljes wrote:
               | Your link's parent page (https://github.com/google)
               | states that they are in the United States of America, on
               | the top, so it's not a good example.
               | 
               | I don't know about your OP, but even as a layperson, I
               | personally like to check where my things come from. And
               | yes, I am mostly curious about which wide geopolitical
               | region the thing is from.
               | 
               | In case of IT projects, it matters when I want to include
               | them in a project.
        
             | spinningarrow wrote:
             | > It's conventional to do so
             | 
             | Where do you see that? e.g. I just checked
             | https://openai.com/about/ and it doesn't say where they are
             | based. I have no associations either way, but I usually
             | have to work hard to find out where startups are based.
        
               | laurentb wrote:
               | it's right there in their terms of use:
               | https://openai.com/policies/terms-of-use/ bottom of each
               | of their pages
        
               | nashashmi wrote:
               | > If you believe that your intellectual property rights
               | have been infringed, please send notice to the address
               | below or fill out this form. We may delete or disable
               | content that we believe violates these Terms or is
               | alleged to be infringing and will terminate accounts of
               | repeat infringers where appropriate.
               | OpenAI, L.L.C.       1455 3rd Street       San Francisco,
               | CA 94158       Attn: General Counsel / Copyright Agent
               | 
               | Is this what you are talking about?
        
             | diggan wrote:
             | > 1. It's conventional to do so.
             | 
             | I can't say I remember any model/weights release including
             | the nation where the authors happen to live or where the
             | company is registered. Usually they include some details
             | about what languages they've included to train on, and
             | disclose some of their relationships, which you could use
             | for inferring that from.
             | 
             | But is it really a convention to include the nation the
             | company happen to be registered in, or where the authors
             | live, in submitted papers? I think that'd stick out more to
             | me, than a paper missing such a detail.
        
               | noelwelsh wrote:
               | OP said "official pages", which I took to mean the
               | company website: https://www.minimax.io/ not the repo or
               | the paper.
        
               | diggan wrote:
               | Ok, lets change the argument to "It's conventional for
               | companies to publish what country they're located in on
               | their project's page", which companies are doing this?
               | Not even OpenAI or Anthropic are doing this as far as I
               | can tell.
        
               | noelwelsh wrote:
               | If you mean a Github page, like text generated from the
               | `README.md`, then I do not expect any mention of country
               | there.
        
               | diggan wrote:
               | > If you mean a Github page
               | 
               | I'm trying to figure out what _you_ mean here. Where do
               | you expect the country to be mentioned?
        
           | npteljes wrote:
           | Forget the project page, I couldn't find definitive
           | information on any of the official pages.
           | 
           | They state HQ in Singapore on LinkedIn, and San Francisco
           | elsewhere. Compared to this, it's outright disingenuous that
           | they don't mention that they are a Chinese company.
           | 
           | As a layman, I'm mostly indifferent to this information.
           | 
           | If I were a project manager, this would be vital information.
           | And the people running projects know this. So it begs the
           | question: why not disclose, and why obscure it?
        
         | 7moritz7 wrote:
         | Many people know that Minimax is Chinese because their video
         | generator has a super obviously Chinese name (Hailuo), and
         | that's what they've been known for so far
        
       | htrp wrote:
       | They apparently building buzz for an IPO
       | 
       | https://www.bloomberg.com/news/articles/2025-06-18/alibaba-b...
        
       | markkitti wrote:
       | Please come up with better names for these models. This sounds
       | like the processor in my Mac Studio.
        
         | chvid wrote:
         | https://en.wikipedia.org/wiki/Minimax
         | 
         | They named themselves after a classic ai algorithm.
        
           | JoeDaDude wrote:
           | As best I can tell from a gloss-over read, it doesn't use
           | anything like the Minimax algorithm. Astute readers are aware
           | that one of the first applications of Minimax was in an AI
           | chess program designed by Claude Shannon.
           | 
           | https://en.wikipedia.org/wiki/Claude_Shannon#Shannon's_compu.
           | ..
        
             | bjord wrote:
             | it's the name of the company
        
             | npteljes wrote:
             | The company supplies contemporary AI solutions, like LLM
             | and video generation. The name is just a reference, like in
             | the case of Tesla, or like how there is a kaliapparat in
             | the American Chemical Society logo.
        
           | badc0ffee wrote:
           | But then there's the "M1" part.
        
         | diggan wrote:
         | Also sounds like my long lost dog whose name was Max but he was
         | tiny. Absolutely horrible name, borderline criminal I say.
        
         | seydor wrote:
         | Your Mac is made by 'Apple' and literally named after an apple
         | cultivar
        
           | kridsdale1 wrote:
           | Is that like a pineapple that doesn't grow on pine trees?
        
       | reedlaw wrote:
       | In case you're wondering what it takes to run it, the answer is
       | 8x H200 141GB [1] which costs $250k [2].
       | 
       | 1. https://github.com/MiniMax-
       | AI/MiniMax-M1/issues/2#issuecomme...
       | 
       | 2. https://www.ebay.com/itm/335830302628
        
         | incomingpain wrote:
         | That's full quantization. If you run Q4 or Q8 you can run this
         | on <$10,000 equipment.
        
           | cma wrote:
           | And if you add in heavy sparsification it should fit and run
           | on a raspberry pi.
        
             | rvz wrote:
             | So in around 6 months, we will see that the person who
             | bought this H200 in the listing just got scammed for $250k
             | and will realize that you just needed specific
             | quantizations to the model and a few optimizations to run
             | locally.
             | 
             | Unless they want to train their own model, buying this for
             | inference for $250k is unnecessary and still isn't enough
             | for a full production deployment.
        
             | vFunct wrote:
             | It's already sparsified from the 150T parameter model..
        
               | yorwba wrote:
               | It took me several hours to realize that 150 trillion
               | parameters is a reference to the number of synapses in a
               | human brain.
        
           | deadbabe wrote:
           | No point in running anything but full quantization.
        
             | esafak wrote:
             | Quantization doesn't work? Really?
        
           | tgtweak wrote:
           | My experience with heavily quantized models is they do better
           | than a similar sized unquantized model but don't really
           | perform anywhere near the pre-quantized model.
        
             | incomingpain wrote:
             | >My experience with heavily quantized models is they do
             | better than a similar sized unquantized model but don't
             | really perform anywhere near the pre-quantized model.
             | 
             | People have tested it. Q8 has essentially no drop in
             | quality, Q4 is measurable but still not realistically a
             | problem. If this impacts you, just pay for the commercial
             | saas option.
        
               | haolez wrote:
               | This assumes that the benchmarks are representative of
               | real usage scenarios. I'm not saying that there is bad
               | faith, but that benchmarking is really hard in the
               | context of LLMs.
        
               | incomingpain wrote:
               | >This assumes that the benchmarks are representative of
               | real usage scenarios. I'm not saying that there is bad
               | faith, but that benchmarking is really hard in the
               | context of LLMs.
               | 
               | It's a fair point, but the conclusion is 'i dont know'
               | 
               | I could assume that it gets better because it'll keep to
               | simpler code.
        
         | GTP wrote:
         | How many parameters does this model have?
        
           | 7moritz7 wrote:
           | 456 bn, about 46 bn active at a time (it's moe)
        
         | kridsdale1 wrote:
         | Can't you run it on a Mac Studio with 512GB? That's about
         | $8,500.
        
       | b0a04gl wrote:
       | if they trained this scale without western cloud infra, i'd want
       | to know what their token throughput setup looks like
        
         | econ wrote:
         | Sneakernet
        
         | jaggs wrote:
         | They trained on 512 H800 GPUs for three weeks, equivalent to
         | around half a million dollars. https://xcancel.com/MiniMax__AI
        
           | yorwba wrote:
           | That is for the reinforcement learning part. The base model
           | was likely trained on more GPUs for significantly longer.
        
       | killerstorm wrote:
       | > "In our attention design, a transformer block with softmax
       | attention follows every seven transnormer blocks (Qin et al.,
       | 2022a) with lightning attention."
       | 
       | Alright, so it's 87.5% linear attention + 12.5% full attention.
       | 
       | TBH I find the terminology around "linear attention" rather
       | confusing.
       | 
       | "Softmax attention" is an information routing mechanism: when
       | token `k` is being computed, it can receive information from
       | tokens 1..k, but it has to be crammed through a channel of a
       | fixed size.
       | 
       | "Linear attention", on the other hand, is just a 'register bank'
       | of a fixed size available to each layer. It's not real attention,
       | it's attention only in the sense it's compatible with layer-at-
       | once computation.
        
       ___________________________________________________________________
       (page generated 2025-06-18 23:00 UTC)