MiMo-V2.5 Has released

[-]

XCSme@reddit

V2.5/V2.5-Pro quite similar to V2-Omni/V2-Pro:

[-]

BananaOrangePie@reddit

I don't trust this leaderboard as it ranks gemini flash above gemini pro and opus...

[-]

XCSme@reddit

The questions are more about raw intelligence than coding capabilities, and Gemini models are definitely the smarted out there.

I was also surprised my Gemini 3 flash (with medium reasoning), but it simply gets everything correct 100% of the time.

On other benchmarks it also is close to the top, or beating latest SOTA models:

[-]

XCSme@reddit

Source: https://aibenchy.com/compare/xiaomi-mimo-v2-5-medium/xiaomi-mimo-v2-5-pro-medium/xiaomi-mimo-v2-omni-medium/xiaomi-mimo-v2-pro-medium/

[-]

9gxa05s8fa8sh@reddit

DON'T BUY HARDWARE, THE AI MARKET IS CRASHING

xaiomi de facto lowering prices while anthropic is de facto raising prices while microsoft is de facto exiting AI by stopping the sale of github copilot because it loses so much money are all MEGA SCARY WARNING SIGNS FOR THE MARKET

[-]

hungy-popinpobopian@reddit

I thought anthropic was running out of compute to keep up with customer demand. Might not be the best time to buy hardware but not convinced its mega scary just yet

[-]

Artistedo@reddit

> "Coming Open Source

> The MiMo-V2.5 series will soon be officially released and open-sourced. Stay tuned.

> Xiaomi MiMo Team · 2026"

I really hope they opensource the pro...

W Xiaomi

[-]

pigeon57434@reddit

pro is basically their version of qwen max its not gonna be open sourced

[-]

Artistedo@reddit

Hmm idk, they actually say "series", I think they would specify if its only one or 2 models, so unless they make a flash and opensource flash, v2.5 normal and tts, I have high hopes for pro

https://mimo.xiaomi.com/mimo-v2-5-pro#:\~:text=The%20MiMo%2DV2.5%20series%20will%20soon%20be%20officially%20released%20and%20open%2Dsourced.

[-]

Illustrious-Many-782@reddit

I think that's what they mean by "series" because 2.5 is the multi modal model and 2.5 pro is text only. So the series would mean more than one model.

[-]

Educational-Formal-8@reddit

i love it...give us fast pls :D

[-]

Cool-Chemical-5629@reddit

I like how everyone is suddenly beating Claude Opus 4.6 when there's Claude Opus 4.7 to compete against.

[-]

Techngro@reddit

Opus 4.7 was only released like a month ago. So if they are actually beating Opus 4.6, it's a big deal.

[-]

john0201@reddit

How big are quants expected to be?

[-]

2muchnet42day@reddit

yes

[-]

Tudragon123456@reddit

At this point I don't need Deepseek V4 anymore :)

[-]

Pink_da_Web@reddit

Yes, you need

[-]

pigeon57434@reddit

i appreciate how they actually care about multimodality i really want high multimodality to become the new standard for oss and more than just vision i mean

[-]

Long_comment_san@reddit

Name is hilarious. In Russian, it means "that missed" lol.

On a serious note, looks incredible

[-]

Dany0@reddit

It means that in other slavic languages too

Anyway I don't see param count anywhere. Mimo v2 only had its "flash" variant get an open weights release, and that one was a whopping 400b params

So my verdict on this model depends a lot on its param count. If it's 1T, it's pathetic (compared to kimi 2.6), only benefit is that it's omni

[-]

PhilippeEiffel@reddit

MiMo-V2 Flash is 309B A15B

[-]

Dany0@reddit

You're right, corrected

[-]

TinyAres@reddit

Their token plan is weak though, so til they fix that at most we can hope for a better alternative provider, that assuming the model is efficient, and based on opencode docs that were updated it could have a place but little sandwhiched between kimi k2.6 and qwen 3.6 / minimax 2.7, and all of these have better other plans too.

https://opencode.ai/docs/go/

https://www.reddit.com/r/opencodeCLI/comments/1sbhdwz/there_goes_my_xiaomi_token_plan/

[-]

9gxa05s8fa8sh@reddit

cost was changed and needs to be re evaluated every new model because different models use different token counts and produce different work

[-]

Old_Stretch_3045@reddit

Relying on my real production experience, I don’t trust Chinese benchmarks (the only ones worth a quick glance are swe-rebench and arc-agi, since they don’t allow benchmaxxing).

[-]