This is a genuinely massive win for open source, and frankly, for the world. MiMo-V2.5-Pro (and even V2 behind it) genuinely feels like Opus at home. So now, for all those people asking "can I run something as good at Claude locally?" the answer is officially technically yes AND you can use it with an MIT license. It's a straight up incredible model, and being able to have access to something this powerful without corporate interference is a game changer for the kind of projects that can be produced by private groups and individuals without reliance on finicky, ethically shady frontier labs. Fuli Luo and her team are absolute legends.
One nice thing about MiMo V2.5 Pro is that it seems to be the most token efficient open source model. It thinks, but doesn't get into a long "actually wait" cycle all the time. Love it.
Is it really that good? I've tried and it wasn't doing that good with some more obscure things like Win32 API calls in a dotnet console app, or using Ghidra via MCP for some reverse engineering tasks reliably.
While claude models were excellent at these. I don't like anthropic and won't pay for their services, but open models have a long way to go in terms of reaching that level of breadth of knowledge
I use GLM-5.1 and Kimi 2.6 for coding plan/build still, but MiMo V2.5 Pro is my favorite model so far for just about everything else. It’s the one I use to figure out what to build, then GLM/Kimi to do the implementation.
This model (and v2 pro before) is incredible, I've never seen anything like it. I'm not sure what is the secret, but it feels like it had no reinforcement to be "likable" by dumb consumer, it's dry and straight to the point most of the times. It even says "I don't know". But it's kinda meh as a coding agent, I use it mainly to discuss tasks or ideas before coding
I'm extremely happy that each lab is diverging in terms of behavior, Kimi, at least K2 (just 2) was also very MiMo-like, hope they keep that. Gemma is much more Claude like, and Qwen... Is Qwen.
I'm surprised at how fast Xiaomi caught up, tbh. It felt like the original model was clowned on, and then one release later they suddenly dominate. Felt like this phenomenon happens with almost all Chinese labs. Maybe that Yuan lab will come next. Hope xiami releases a smaller (maybe 30-80B) model soon for the GPU poor, though
At this point the Chinese are just rubbing their dicks on Silicon Valley's face. They don't have a SOTA AI model, they have like 10. And all free. And every week they publish new one.
These models aren't better than Anthropic or OpenAi's models. Yes, they're free and open weight which is a middle finger to them (and amazing for us) but your comment reads like these Chinese models far surpass the West's products but that's really not true
But that's not how business thinks. It's a weighted model. If the raw intelligence and price are equal to a business then these models do "win". If my business were AI dependent the raw intelligence gap would need to be far, far larger for my to consider paying for tokens at this point.
You still have to pay for hosting and depending on how many tokens you use, its likely far cheaper to use openai/anthropic instead of hosting the model on your own
Even if I agreed (I don't, we host both Kimi and Mini, largely due to sensitivity, and it's worth the costs) that statement assumes there won't be massive price hikes. Given the money OpenAI needs to raise (nearly $1 trillion) I think people paying monthly will be experiencing sticker shock soon.
I think it would be the opposite. Prices will likely come down due to competition, inference optimization, hardware development, architecture progress.
I wonder if a bit of both might happen. As in, the price relative to the intelligence levels of the models will probably continue to improve over time, but on the other hand, the raw price (ignoring intelligence level) will probably go up over time.
So, higher subscription prices or price per token over time, but also significantly stronger models (with the rate of strength increasing faster than the rate of the price hikes) over time.
Somebody else did the math and it costs anthropic something like $5000 to service the $20/mo plans.
Inference costs are falling quickly but eventually investors are going to want to get their investment back so id guess maybe $2000/mo for a "software developer replacement" amount of tokens.
I do wonder what the US will do besides "banning Chinese models". Even closing the gap is an existential threat to the big companies and you can't have all the dominant models produced by a single country.
they are going to restrict access to higher models to their inside circles of companies, limiting consumer end-user from direct access, pushing through various vendors i guess
they are behind for very little given the timelapse between startups and the cost difference make it worth.
plus we are almost at the finish line of what large language models are capable, so if they get to offer the same quality, who is going to pay 10x the cost ?
They don't have to be better, they just have to be acceptable to set a ceiling on the profitability of AI inference providers. I personally use open models that are hosted at inference providers for all of my enterprise workflows. Not all of my employees do, but right now I don't care. In the future I probably will.
Can you imagine how much better Chinese models woud be if they are able to buy the latest EUV systems or even just the Nvidia chips without worrying about sanctions? EUV is basically the combined efforts of US, EU, Japan and SK and China has to tackle it all by itself.
although you are not wrong, you miss the main point. The models aren't better, but they are close, and they are definitely good enough for the big majority of cases. More, they are catching up, the gap keeps closing and getting narrower every month
I used all the models on large and complicated code bases and no they are not SOTA. They can compete because of the price but GPT5.5 blows them out of the water in performance
They're trying to collapse the AI market and given the attack on American workers by the country's own developers it's hard for me to not cheer for them a bit.
For many companies like google, meta, xiaomi who make plenty of money in other areas, it makes business sense. Their moat is they can afford to not make much money from it and have it commoditized. The ai super labs cannot.
I don't think they are saints or heroes, once they succeed and collapse the competition, they will back at charging lots of money. They do this only because make business sense.
Its cool because unless the US restricts access to Chinese models, Anthropic and Open AI will have a hard time attracting customers to their paid plans.
I believe this work as a hard limit on how much they can charge. Weren't for chinese models, OpenAI could easily charge thousands for their regular models.
OpenAI could easily charge thousands for their regular models.
The existence of Chinese models is not what keeps the plans so cheap right now. They're trying to rapidly gain marketshare. Chinese models or not it was always going to be cheap. Initially...
Due cheap powers and cheaper labor, plus goverment incentivies, the initial cost is already way cheaper to Western rivals. You need to take this into account as well...
I think the vast majority of people who use AI (I don't mean power users) have no idea that a Chinese model even exists. Or even what a model is for that matter.
In my case, my company knows what they are but still thinks Chinese = Bad.
People would think the same before opus 4.5. Look the current state. Masses move a little behind, but since the masses here are devs, they don’t move so slowly or have fidelity. Everyone will make the switch eventually.
But you can fine tune it, re-training from scratch would cost hundreds of thousands of dollars at least, just to re-create the same result again.
It is a huge leap to be able to spare that and run only a couple of thousands maybe and upgrade it for your specific nieche of work, run it on your stack as you please.
Now do this with ChatGpt : you have oss20b and 120b which are great for a.. customer support job ?
Even if they did that, Europe and other first world countries would use the on par, cheaper, Chinese services over the absurd prices the American providers demands. Maybe even host their own.
Active params are used to evaluate performance, not for VRAM usage, RAM offloading to putting almost all weights in RAM, VRAM is mainly for kvcache, so techinically you don't even need 2 cards.
Yann Lecun gave out some amazing talks available online.
And it's very clear dude has never bothered to change his ppt template for years. It even looks like he's just editing texts and replacing images on the same ppt file.
me waiting a 100b a15b or 200b a24b or a similar model which never comes... it is either 35b/27b or 1t a50b model. They are trying hard to force 5090 rig users to 6000 pros
I know it probably isn’t but it would be so funny if one guy was uploading the weights and running the Twitter page from the same machine frantically tabbing between the ssh session and the browser to live tweet it
These new MiMo V2.5 (both pro and non-pro) perform extremely strongly in my own benchmarks and are heavily underrated. This is truly a huge gain for the open-source community!
im surprised they open sourced the pro version i just though they would do regular 2.5 thats very nice
and it seems this may be the new most capable oss base in the world and possibly most capable model period but K2.6 might still edge it out for me but it still uses K2-base so you cant build off it as easily if thats something you do
Just made an linux server app with GUI with mimo pro2 and opencode. Spent 8 euros in code- fast, smart, fast debug. Happy, maybe i will try deepseek, is even less expensive.
This is what I have been waiting for! I love this model on API its debatably better then K2.6 across my testing, Tad bit worse at coding but 1M context and low hallucination rates make it nicer to use in almost all aspects!
seamonn@reddit
I wish this was multi modal
Ok_Technology_5962@reddit
But the flash...
LoveMind_AI@reddit
This is a genuinely massive win for open source, and frankly, for the world. MiMo-V2.5-Pro (and even V2 behind it) genuinely feels like Opus at home. So now, for all those people asking "can I run something as good at Claude locally?" the answer is officially technically yes AND you can use it with an MIT license. It's a straight up incredible model, and being able to have access to something this powerful without corporate interference is a game changer for the kind of projects that can be produced by private groups and individuals without reliance on finicky, ethically shady frontier labs. Fuli Luo and her team are absolute legends.
ebra95@reddit
I have not used Opus, I have used many others and MiMo-V2.5-Pro is already more than enough for what most people can do with it.
It's awesome.
Accomplished-Air439@reddit
One nice thing about MiMo V2.5 Pro is that it seems to be the most token efficient open source model. It thinks, but doesn't get into a long "actually wait" cycle all the time. Love it.
ebra95@reddit
Yes indeed it's fast for a 1T parameter. I guess the infrastructure holds up well.
power97992@reddit
Maybe like opus 4.5 not 4.6/4.7 or gpt 5.5
Icy_Butterscotch6661@reddit
Is it really that good? I've tried and it wasn't doing that good with some more obscure things like Win32 API calls in a dotnet console app, or using Ghidra via MCP for some reverse engineering tasks reliably.
While claude models were excellent at these. I don't like anthropic and won't pay for their services, but open models have a long way to go in terms of reaching that level of breadth of knowledge
TheReedemer69@reddit
Lets connect.
real_serviceloom@reddit
Yup it is really good with rust in my case. Better than Kimi 2.6 and GLM 5.1.
look@reddit
I use GLM-5.1 and Kimi 2.6 for coding plan/build still, but MiMo V2.5 Pro is my favorite model so far for just about everything else. It’s the one I use to figure out what to build, then GLM/Kimi to do the implementation.
drumyum@reddit
This model (and v2 pro before) is incredible, I've never seen anything like it. I'm not sure what is the secret, but it feels like it had no reinforcement to be "likable" by dumb consumer, it's dry and straight to the point most of the times. It even says "I don't know". But it's kinda meh as a coding agent, I use it mainly to discuss tasks or ideas before coding
look@reddit
Exactly. MiMo V2 Pro is the model I talk to and work through ideas with, and everything else is effectively just a subagent to it.
ComplexType568@reddit
I'm extremely happy that each lab is diverging in terms of behavior, Kimi, at least K2 (just 2) was also very MiMo-like, hope they keep that. Gemma is much more Claude like, and Qwen... Is Qwen.
I'm surprised at how fast Xiaomi caught up, tbh. It felt like the original model was clowned on, and then one release later they suddenly dominate. Felt like this phenomenon happens with almost all Chinese labs. Maybe that Yuan lab will come next. Hope xiami releases a smaller (maybe 30-80B) model soon for the GPU poor, though
jnmi235@reddit
The non-pro model looks strong but needs 4x rtx pro cards in their mixed precision
silenceimpaired@reddit
I'm excited to try out MiMo-V2-Flash Quant 0.01km tomorrow ;)
ComplexType568@reddit
Heard the 0 bit quant can run on a pi
ortegaalfredo@reddit
At this point the Chinese are just rubbing their dicks on Silicon Valley's face. They don't have a SOTA AI model, they have like 10. And all free. And every week they publish new one.
smith7018@reddit
These models aren't better than Anthropic or OpenAi's models. Yes, they're free and open weight which is a middle finger to them (and amazing for us) but your comment reads like these Chinese models far surpass the West's products but that's really not true
jld1532@reddit
But that's not how business thinks. It's a weighted model. If the raw intelligence and price are equal to a business then these models do "win". If my business were AI dependent the raw intelligence gap would need to be far, far larger for my to consider paying for tokens at this point.
KeikakuAccelerator@reddit
You still have to pay for hosting and depending on how many tokens you use, its likely far cheaper to use openai/anthropic instead of hosting the model on your own
jld1532@reddit
Even if I agreed (I don't, we host both Kimi and Mini, largely due to sensitivity, and it's worth the costs) that statement assumes there won't be massive price hikes. Given the money OpenAI needs to raise (nearly $1 trillion) I think people paying monthly will be experiencing sticker shock soon.
KeikakuAccelerator@reddit
I think it would be the opposite. Prices will likely come down due to competition, inference optimization, hardware development, architecture progress.
DeepOrangeSky@reddit
I wonder if a bit of both might happen. As in, the price relative to the intelligence levels of the models will probably continue to improve over time, but on the other hand, the raw price (ignoring intelligence level) will probably go up over time.
So, higher subscription prices or price per token over time, but also significantly stronger models (with the rate of strength increasing faster than the rate of the price hikes) over time.
Well, who knows, but that's my guess, for now.
jld1532@reddit
I guess we'll see. However, I would point to token based billing as the first price increase just in the form of artificial scarcity.
KeikakuAccelerator@reddit
It is not really artificial scarcity. Demand is higher than supply.
StatusSociety2196@reddit
Somebody else did the math and it costs anthropic something like $5000 to service the $20/mo plans.
Inference costs are falling quickly but eventually investors are going to want to get their investment back so id guess maybe $2000/mo for a "software developer replacement" amount of tokens.
Meanwhile, qwen 3.6 is free...
rhythmdev@reddit
hell yes.
NoahFect@reddit
Reddit fuzzes votes
Acu17y@reddit
They are
TechSwag@reddit
You gotta be a bot
or illiterate.
Acu17y@reddit
Have you ever tried kimi 2.6 ? It costs 1/20 of the Opus 4.7 and is on par
ortegaalfredo@reddit
If you think it's on par, then you don't give it hard enough problems. Opus is far better but then again, not everybody needs it.
Acu17y@reddit
It is, and my project is very big. I don't take money from anyone to say it.
Dabalam@reddit
I do wonder what the US will do besides "banning Chinese models". Even closing the gap is an existential threat to the big companies and you can't have all the dominant models produced by a single country.
ebra95@reddit
they are going to restrict access to higher models to their inside circles of companies, limiting consumer end-user from direct access, pushing through various vendors i guess
ebra95@reddit
they are behind for very little given the timelapse between startups and the cost difference make it worth.
plus we are almost at the finish line of what large language models are capable, so if they get to offer the same quality, who is going to pay 10x the cost ?
RedParaglider@reddit
They don't have to be better, they just have to be acceptable to set a ceiling on the profitability of AI inference providers. I personally use open models that are hosted at inference providers for all of my enterprise workflows. Not all of my employees do, but right now I don't care. In the future I probably will.
Jackw78@reddit
Can you imagine how much better Chinese models woud be if they are able to buy the latest EUV systems or even just the Nvidia chips without worrying about sanctions? EUV is basically the combined efforts of US, EU, Japan and SK and China has to tackle it all by itself.
ortegaalfredo@reddit
It's true, but they are at max, 2 months behind them. Thats not enough of a moat.
nunodonato@reddit
although you are not wrong, you miss the main point. The models aren't better, but they are close, and they are definitely good enough for the big majority of cases. More, they are catching up, the gap keeps closing and getting narrower every month
Endoky@reddit
I used all the models on large and complicated code bases and no they are not SOTA. They can compete because of the price but GPT5.5 blows them out of the water in performance
rhythmdev@reddit
5.5 is what... 10 days old? Give it a couple months.
jld1532@reddit
They're trying to collapse the AI market and given the attack on American workers by the country's own developers it's hard for me to not cheer for them a bit.
Monkey_1505@reddit
For many companies like google, meta, xiaomi who make plenty of money in other areas, it makes business sense. Their moat is they can afford to not make much money from it and have it commoditized. The ai super labs cannot.
ortegaalfredo@reddit
I don't think they are saints or heroes, once they succeed and collapse the competition, they will back at charging lots of money. They do this only because make business sense.
throw_me_away3478@reddit
Its cool because unless the US restricts access to Chinese models, Anthropic and Open AI will have a hard time attracting customers to their paid plans.
Another win for open source models
Chris266@reddit
I wouldnt say Anthropic and OpenAI are having a hard time attracting people to their plans...
ortegaalfredo@reddit
I believe this work as a hard limit on how much they can charge. Weren't for chinese models, OpenAI could easily charge thousands for their regular models.
xienze@reddit
The existence of Chinese models is not what keeps the plans so cheap right now. They're trying to rapidly gain marketshare. Chinese models or not it was always going to be cheap. Initially...
Um0therfckers@reddit
Due cheap powers and cheaper labor, plus goverment incentivies, the initial cost is already way cheaper to Western rivals. You need to take this into account as well...
Chris266@reddit
I think the vast majority of people who use AI (I don't mean power users) have no idea that a Chinese model even exists. Or even what a model is for that matter.
In my case, my company knows what they are but still thinks Chinese = Bad.
jld1532@reddit
Universities are using open weights models. These ecosystems are actively being incorporated into higher education. It's only a matter of time.
mestresamba@reddit
People would think the same before opus 4.5. Look the current state. Masses move a little behind, but since the masses here are devs, they don’t move so slowly or have fidelity. Everyone will make the switch eventually.
xrvz@reddit
Something's bound to implode and such companies will learn the hard way.
PinkySwearNotABot@reddit
You mean another win for us. Think about what America would do to us consumers if China and their open models weren’t there to give us leverage.
ortegaalfredo@reddit
Its not really open-source, more like open-weights, and you cannot truly recreate the tech, just use it for free.
ebra95@reddit
But you can fine tune it, re-training from scratch would cost hundreds of thousands of dollars at least, just to re-create the same result again.
It is a huge leap to be able to spare that and run only a couple of thousands maybe and upgrade it for your specific nieche of work, run it on your stack as you please.
Now do this with ChatGpt : you have oss20b and 120b which are great for a.. customer support job ?
kevinlch@reddit
most people don't care as long as it's free
bbjurn@reddit
Still much better than what most US companies are providing.
More-Curious816@reddit
Even if they did that, Europe and other first world countries would use the on par, cheaper, Chinese services over the absurd prices the American providers demands. Maybe even host their own.
RedParaglider@reddit
What it's really doing is basically saying that the cap on what inference providers can charge is not based on hardware+markup, it's just hardware.
Budget-Juggernaut-68@reddit
It's actually working pretty well based on deepseek V4's recent report.
rageling@reddit
They don't have like 10, they have like 0, GPT 5.5 and Opus 4.7 are SOTA, China is offering nothing of comparable quality
ortegaalfredo@reddit
Those models are 2 weeks old. They have about 2-3 month advantage, no more.
Icy_Butterscotch6661@reddit
You think you are talking to someone who woke up a loser?!
Turbulent_Pin7635@reddit
AI soon will be like ping pong, it is easier to win the world championship than win the Chinese championship
ComplexType568@reddit
oh my goodness another 1T baby. i want to see how GLM 5.X, Kimi K2.X (or 3) and DeepSeek V4 all play out...
ilintar@reddit
Gonna try it out as soon as my cat gives back the stack of 8 RTX 6000s he's hidden somewhere.
pmttyji@reddit
Attract the kitty with 8 Treats
srigi@reddit
Cat attracts him with 8x RTXs
ChocomelP@reddit
Get pussy with CUDA?
Weak_Kaleidoscope839@reddit
And a pspspspspsps
UpAndDownArrows@reddit
I mean come on people. It's a MoE with 49B active params. A 5090 paired with 3090 can fit the active experts, and the rest can live in RAM.
ilintar@reddit
Yeah, cat also ate 1GB worth of DDR5 chips, mistook them for dry food.
coder543@reddit
1 gigabyte? In this economy?
ilintar@reddit
Sorry, meant 1TB obviously, cat only eats 64+ GB DDR5 wafers, he's pretty picky.
blackbird2150@reddit
64gb, nothing less!
czktcx@reddit
Active params are used to evaluate performance, not for VRAM usage, RAM offloading to putting almost all weights in RAM, VRAM is mainly for kvcache, so techinically you don't even need 2 cards.
But where's my 512G RAM?
UpAndDownArrows@reddit
I mean I get it, 512 GB RAM is not laying on a sidewalk, but like...
Compare getting 512 GB RAM vs getting 8 RTX 6000s, the difference is just an order of magnitude or more.
grumd@reddit
Wake me up when V2.5-Flash drops
__Maximum__@reddit
You meant the REAP-ed 1bit quants of the flash
ambient_temp_xeno@reddit
Good news, bad news situation:
https://huggingface.co/XiaomiMiMo/MiMo-V2.5
pkmxtw@reddit
Why are those labs capable of training multi million dollar models and yet are so terrible at making charts lol
smallDeltaBigEffect@reddit
you should read some papers of the smartest people currently alive.
If it wasnt for some design affine phd students, you would get handdrawn powerpoint slides
zdy132@reddit
Yann Lecun gave out some amazing talks available online.
And it's very clear dude has never bothered to change his ppt template for years. It even looks like he's just editing texts and replacing images on the same ppt file.
Dany0@reddit
That skill is not transferrable. The could train a model to make charts for them, but they cba
a_beautiful_rhind@reddit
That's the one because I didn't buy enough ram, lol.
segmond@reddit
Apple needs to hurry the fuck up and let us know the spec for m5 ultra stdio so we can decide if we are buying it or a horde of Nvidia GPUs.
power97992@reddit
It should be already by end of may or june
rhythmdev@reddit
me waiting a 100b a15b or 200b a24b or a similar model which never comes... it is either 35b/27b or 1t a50b model. They are trying hard to force 5090 rig users to 6000 pros
SnooPaintings8639@reddit
But the smaller version is 300b A15, so close enough. Go lower with a quant and have fun. With only 15b active params it should be usable.
rhythmdev@reddit
Y either that or deepseek flash, i'll try my chances with those but i dont have high hopes. Or wait for a distilled smaller version
GreedyWorking1499@reddit
What is “F8_E4M3”?
fantasticsid@reddit
An FP8 format with 4 exponent bits, 3 mantissa bits, and a sign bit.
GreedyWorking1499@reddit
Is that how all FP8 is?
fantasticsid@reddit
Nah, there's also E5M2 - which trades range for precision - and a bunch of specialised stuff like MXFP8 (same trick as MXFP4, just twice as wide.)
GreedyWorking1499@reddit
How does that work if a weight is big enough to be stored in E5M2 but too big for E4M3? What is the range of weights in actual GGUFs?
fantasticsid@reddit
No idea tbh - I assume the weights in question don't exceed the range of E4M3 so they used that format for the extra precision it provides?
GreedyWorking1499@reddit
Interesting. I wonder why people would use E5M2 then
SnooPaintings8639@reddit
Image, video, audio, long context... the smaller version of 310B A15 might end up being a MiniMax 2.5 replacement.
ThePixelHunter@reddit
Should be about 140B in Q3. Can't wait to run it.
Dramatic-Rub-7654@reddit
1-bit gguf when?
LegacyRemaster@reddit
Another 2x rtx 6000 needed
RedParaglider@reddit
YOU MUST CONSTRUCT ADDITIONAL PYLONS
onewheeldoin200@reddit
😂
datbackup@reddit
WE REQUIRE MORE VIDEO RAM
mindwip@reddit
Omg yes
LegacyRemaster@reddit
for Adune
Dany0@reddit
Why do I get spam enticing me with women. Entice me with H200s, damn it!
Eyelbee@reddit
More like 16x
LegacyRemaster@reddit
q4 :D
SeaDisk6624@reddit
8 for nvfp4
LegacyRemaster@reddit
yes but I have already 96+48+48gb vram + 128gb ram :D
lendo93@reddit
Most underrated model of this release cycle. We have it performing better than DeepSeek and comparable to Kimi 2.6, without the rate limits.
nullmove@reddit
Can't run either. At least this was funny: https://xcancel.com/XiaomiMiMo/status/2048803550562844727#m
adumdumonreddit@reddit
I know it probably isn’t but it would be so funny if one guy was uploading the weights and running the Twitter page from the same machine frantically tabbing between the ssh session and the browser to live tweet it
RnRau@reddit
They used QAT for the fp8 weights?
I guess this snippet from the model card suggest that is the case - "Trained on 27T tokens using FP8 mixed precision"
jake_schurch@reddit
Fantastic that my hardware is only off by 1 order of magnitude instead of multiple
kyleboddy@reddit
Mimo-v2.5-Pro benchmarks extremely well on biomech-bench that we launched today though it consumes a ton of tokens comparatively.
https://x.com/drivelinekyle/status/2048775424621396472
Zestyclose-Ad-6147@reddit
Whoo! 🙌
jzn21@reddit
These new MiMo V2.5 (both pro and non-pro) perform extremely strongly in my own benchmarks and are heavily underrated. This is truly a huge gain for the open-source community!
funding__secured@reddit
Can't run on my RTX 6000 Pros yet 😞 Needs FlashAttention 3.
vinigrae@reddit
No way they open sourced that
JC1DA@reddit
Is this better than Qwen-3.5-397B, it's smaller but it lacks of vision capability
pigeon57434@reddit
it does have vision? what the fuck are you talking about it also has audio input too
pigeon57434@reddit
im surprised they open sourced the pro version i just though they would do regular 2.5 thats very nice
and it seems this may be the new most capable oss base in the world and possibly most capable model period but K2.6 might still edge it out for me but it still uses K2-base so you cant build off it as easily if thats something you do
maxpayne07@reddit
Just made an linux server app with GUI with mimo pro2 and opencode. Spent 8 euros in code- fast, smart, fast debug. Happy, maybe i will try deepseek, is even less expensive.
_hephaestus@reddit
How many mac studios do I need to run this
myreala@reddit
Just one 512gb would work for the v.2.5, maybe 3 for pro version.
Kiedrola@reddit
Larga vida a China! Gracias!
jochenboele@reddit
Already have been testing the pro version for a few days. Can’t wait until they opensource it
coder543@reddit
What do you think this post is that you're commenting on?
jochenboele@reddit
😂😂 I saw the model name and jumped to conclusions. I searched for the opensource this morning, but it wasn’t there yet 😂
SeaDisk6624@reddit
the non pro will be great with 4x 6000
Long_comment_san@reddit
*laughs maniacally*
ghgi_@reddit
This is what I have been waiting for! I love this model on API its debatably better then K2.6 across my testing, Tad bit worse at coding but 1M context and low hallucination rates make it nicer to use in almost all aspects!
jacek2023@reddit
too big for my 72/84GB of VRAM
j_osb@reddit
MIT licenses is not bad