TheaterFire

Apparently Minimax 2.7 will be closed weights

Posted by tarruda@reddit | LocalLLaMA | View on Reddit | 54 comments

Reply to Post

54 Comments

Goldandsilverape99@reddit

https://preview.redd.it/84wuluda5dqg1.png?width=1705&format=png&auto=webp&s=59d97f329b41729c2e2e0a9551dad85dd4426b87 Proprietary according to artificialanalysis
View on Reddit #81219794

Exciting-Engine882@reddit

the guys/girls at xiaomi ai, stated on X that they will release the weights for mimo v2 pro
View on Reddit #81266379

TechnoByte_@reddit

It's *currently* proprietary, because they haven't released the weights *yet*. However this doesn't mean they're never coming and artificial analysis is not Minimax, so not a reliable source on this
View on Reddit #81255775

hainesk@reddit

M2.5 took awhile to get to Huggingface too. Is there a source from Minimax that states that 2.7 will be closed source?
View on Reddit #81204267

TechnoByte_@reddit

Nope, no official source, just twitter users speculating
View on Reddit #81255598

_seelos@reddit

https://preview.redd.it/f8x641fhqeqg1.png?width=698&format=png&auto=webp&s=4b27fd40307b3437239a382f718ab37c51270cc5 OpenCode's Zen API (which provides an inference gateway on US servers for opensource models) already has a git commit prepared to launch Minimax 2.7. So it looks like it's going to be open weights.
View on Reddit #81234407

tarruda@reddit (OP)

What is your source about OpenCode Zen API being a gateway only for open weight models? I've looked at https://opencode.ai/docs/zen/ and didn't find anything. In fact, it shows many proprietary models such as GPT being part of the gateway.
View on Reddit #81235773

_seelos@reddit

Yea I'll add some details. They're going to offer the model it in their 'Go' subscription which is only OSS models. https://preview.redd.it/hijgq3ojveqg1.png?width=1771&format=png&auto=webp&s=3b0f5f075eb2bdc4b7417c3c971363c021f7c945
View on Reddit #81236062

tarruda@reddit (OP)

Seems like a good evidence. Thanks for sharing!
View on Reddit #81238026

jacek2023@reddit

Does it really matter to people who only pretend to use AI locally? I use Minimax locally but I am tiny minority here. Most of you use cloud and hype benchmarks.
View on Reddit #81235215

FullOf_Bad_Ideas@reddit

This source is mostly speculative though silence does suggest that there's a chance it will be closed. In that case I wonder if someone will distill M2.7 onto M2.5. I think minimax has a chance to get a lot of revenue and even moat from OpenClaw craze. They are public now, so they'll do what's the best for shareholders.
View on Reddit #81200664

Exciting_Garden2535@reddit

>They are public now, so they'll do what's the best for shareholders. 1. Does that mean, if you are private, you don't do what's best for shareholders? 2. Millions, or perhaps billions of people are shareholders of Google, me included: I have some pieces of their stock because I have some S&P500. Please tell me how I can command them to do something? :)
View on Reddit #81203693

xXprayerwarrior69Xx@reddit

When you are private, the owners may be smart enough to understand the concept of delayed rewards
View on Reddit #81228164

FullOf_Bad_Ideas@reddit

>Does that mean, if you are private, you don't do what's best for shareholders? If you are private, your shareholders might be different, which may allow you to grow the company in a different way. >Millions, or perhaps billions of people are shareholders of Google, me included: I have some pieces of their stock because I have some S&P500. Please tell me how I can command them to do something? :) expectation of growth with bigger investors pushing for changes - https://www.cnbc.com/2022/11/15/alphabet-must-cut-headcount-and-trim-costs-activist-investor-tci-says.html
View on Reddit #81216434

viperx7@reddit

i just hope that they didn't just change their mind after witnessing the kimi k2.5 rebranding by cursor. it makes sense why spend money and resource when someone is going to steal your work and call it as your own
View on Reddit #81226324

twack3r@reddit

I hope not. The cadence between cloud and huggingface release was also a topic of discussion for M2.5 but every day those weights don’t show up point towards a closed release. I hope not but it’s not looking unlikely.
View on Reddit #81224185

xAragon_@reddit

What's the point? Who would use it when there are models like Claude / GPT / Gemini that are likely to be much better?
View on Reddit #81199990

cheechw@reddit

Way better rate limits, like easy 20-100x as much for the same cost.
View on Reddit #81201845

xAragon_@reddit

If you're using the API with OpenRouter - there are no rate limits. As for the cost (which I think you're confusing with rate limits?) - a model like Gemini Flash 3 is likely to be better and perhaps even faster at a similar price (of course I'm just assuming as we can't know for sure).
View on Reddit #81219303

SucculentSpine@reddit

Open sourcing in a capital heavy investment industry is just not viable long-term, especially when there's so much competition. These companies literally have no choice but to try and start making money. They can't continue to get investment otherwise.
View on Reddit #81201566

FullOf_Bad_Ideas@reddit

It's a cheap model hosted on API for openclaw, lots of people are waiting for the moment they'll be able to stop using Sonnet/Opus there. So, if they go closed here, it's because they want to capture this inference revenue and maybe also more Openclaw training data to build out a moat.
View on Reddit #81200424

maglat@reddit

Winter time for open weights. makes me super sad…
View on Reddit #81196122

fredandlunchbox@reddit

Nvidia going strong. Nemotron is very good.
View on Reddit #81214769

Due-Memory-6957@reddit

Because of one company? Common.
View on Reddit #81207037

coder543@reddit

This X post is referencing the same crappy VentureBeat article that everyone else has pointed to. There is no reason to believe that VentureBeat actually knew anymore than anyone else. I am worried that MiniMax M2.7 will be closed, but we still have no proof one way or another.
View on Reddit #81202487

Aaaaaaaaaeeeee@reddit

Surely you remember some lemon saying 2.5 would be closed too?  But they've gone ahead and said they might consider NVFP4 on M3. They don't talk like they're stopping open weights? 
View on Reddit #81198425

HopePupal@reddit

is this a real source or just some jerkoff Twitter user speculating
View on Reddit #81202084

Betadoggo_@reddit

The second one
View on Reddit #81214745

Orlandocollins@reddit

Its just post training on the previous weights to gain a bump. I think they would keep the trend orf releasing given that
View on Reddit #81214178

GCoderDCoder@reddit

The people who would use the open weight versions aren't about to pay for the api anyway so they're just killing their community position by closing weights.
View on Reddit #81212664

Conscious_Cut_6144@reddit

It's just some rando advertising his service. Open weights is coming be patient.
View on Reddit #81210857

__JockY__@reddit

Why are we treating randos on Twitter as primary sources? He's just shilling for Novita.
View on Reddit #81206396

Exciting_Garden2535@reddit

Just the opinion of some people on the Internet. Do they work in MiniMax, or maybe speak with them about this matter? On the M2.7 page (https://www.minimax.io/news/minimax-m27-en), MiniMax stated: >2. We have also enhanced the model's expertise and task delivery capabilities across various fields in the professional office software domain. Its ELO score on GDPval-AA is 1495, the highest among open-source models. Why compare with open-source models if you are not going to open the model weights?
View on Reddit #81203298

kingo86@reddit

That's what I'm clinging onto too - I really hope it's true, but it's not bulletproof. "Best among open source" sounds a lot better than "nearly as good as frontier models". Given their track record producing open weights, maybe they thought they could slip this past users who thought it could logically be open.
View on Reddit #81205815

twavisdegwet@reddit

It'll be released as open weight. You can cite me. I have at least much pull as these random tweets
View on Reddit #81204034

TokenRingAI@reddit

Me too, I will put my non-existant reputation on record for this one.
View on Reddit #81205367

Rascazzione@reddit

The same thing happened with MiniMax 2.5; people were afraid it would be closed off again. But to this day, MiniMax stands by its statement: make affordable intelligence for everyone. If we are using their models, the least we can do is pay for a subscription and buy as many shares as possible.
View on Reddit #81205337

inevitabledeath3@reddit

Guys can we just ignore twitter users speculating? They don't always release weights right away, this lab has had a delay in doing that before.
View on Reddit #81205142

KvAk_AKPlaysYT@reddit

https://preview.redd.it/14nf2a98vaqg1.png?width=1440&format=png&auto=webp&s=7dc9c34df4ab3ab2ff0600e44ee0299de2a15f70 Man... They baited us :(
View on Reddit #81203916

LoveMind_AI@reddit

Well, that's absolutely lame. It's a great model - I like it a lot better than 2.5 - but without the ability to fine-tune it, I wouldn't go near it for production.
View on Reddit #81202109

tarruda@reddit (OP)

Additional sources: https://x.com/0xSero/status/2035078557538320584 https://venturebeat.com/technology/new-minimax-m2-7-proprietary-ai-model-is-self-evolving-and-can-perform-30-50
View on Reddit #81195615

illiteratecop@reddit

All of these source back to this venturebeat article, which as far as I can tell just calls it proprietary because the weights aren't out yet. Zero information here until someone actually affiliated with Minimax says something one way or another.
View on Reddit #81197572

Zc5Gwu@reddit

NovitaAI on openrouter is currently hosting it as well... not sure what to make of that.
View on Reddit #81202067

Tr4sHCr4fT@reddit

Imagine everyone got some days of after the announcement, to reconcile with family or something, and they return on Monday to this.
View on Reddit #81201308

Zc5Gwu@reddit

Bummer. I was mainly using it for the good will. Why would I use it over gemini flash which is roughly the same price?
View on Reddit #81197083

RedParaglider@reddit

Because it's no cap better than Flash?
View on Reddit #81201783

Snoo_64233@reddit

Use it because you need to pragmatically, not because you have some religious dogma. It is okay to pivot away.
View on Reddit #81200991

Zc5Gwu@reddit

If I want to support open source and good behavior of corporations, I don't see why I shouldn't vote with my wallet.
View on Reddit #81201733

FullOf_Bad_Ideas@reddit

It's not the same price at all. Previous Gemini Flash models were really cheap but Gemini 3 Flash is about 2x more expensive than Minimax M2.5, even if you include prompt caching in the calculation.
View on Reddit #81200325

rorowhat@reddit

They're done if they do...people will move to open every time.
View on Reddit #81198319

RedParaglider@reddit

I don't know I see a lot of people using Opus and gipity including me.
View on Reddit #81201746

SteppenAxolotl@reddit

Why would they care? Do you think they are doing this for status. All the western free riders aren't exactly paying their bills.
View on Reddit #81199127

phenotype001@reddit

I won't even test it.
View on Reddit #81200025

NoFaithlessness951@reddit

https://preview.redd.it/v9gglua8aaqg1.jpeg?width=600&format=pjpg&auto=webp&s=8cff6ad842316485fd86b25253fa45f7d9a11ed5
View on Reddit #81197764