Last time, we had the Chinese store-dash company called Meituan join the AI race. But this time, we also have Xiaomi, the Chinese smartphone company,
joining the AI race, too. And what's even more incredible is that within 1 year since their first ever LLM release, they have basically topped the
open-source leaderboard just like that. So, what is going on? How can a company not only successfully
get into the phone business, the home appliance business, the EV business, and now the LLM business? So, in today's video, let us break down the
legendary run Xiaomi is currently on by diving into their research, their models, and find out the reason why they have been able to catch up in such
a short amount of time. But before we dive into it, with new AI models popping out one after the other,
subscribing to every single one of them just to test them out can get expensive pretty quickly. Because realistically, why pay for five different
subscriptions when you can access a full bundle of models in one place for just 10 bucks a month? And this is where Mammut AI comes in. They give you
access to models like Claude, GPT, Gemini, Llama, Mistral, Grok, DeepSeek, Perplexity, Nano, Banana, and
Recraft all in one platform. So, whether you want to test different LLMs, generate images, or compare which model gives you the best result, you can
do it without constantly switching between a bunch of different tabs or subscription services. And what makes it even more useful is that Mammut also
lets you reprompt, compare answers side by side, and even have models challenge each other so you can
quickly choose the strongest output instead of manually checking everything yourself. It also supports a multi-model workflow where you can analyze
documents and images, use Deep Research through Perplexity, and even use voice chat and dictation when you want to move faster. As for repetitive
tasks, you can use your project function to create your own custom Mammuts with your own instructions,
preferred models, and organized workflows so you don't have to start from scratch every time. So, if you want one clean place to use some of the best
AI models without spending money on a bunch of separate subscriptions, check out Mammut AI using link down in description, and thank you Mammut for
sponsoring this video. Anyways, even though our first impression of Xiaomi is it being a phone
company, when it was first founded back in 2010, their first product was actually not a phone. Their first gig was something called the MIUI, which is
a now deprecated mobile operating system based on Android launched in August 2010. All because this guy, Lei Jun, thought all the current phones sucks
and he could make a better one. So, after they developed the system and heavily iterated it, they
secured a 41 million Series A and within only 1 year, they released their first smartphone, the Xiaomi Mi 1. Their strategy was simply building an
operating that people liked, then build hardware around that software, but most importantly, sell it at an insanely aggressive price. So, thanks to
the size of China, the internet, the supply chain execution, and all the feedback they got, they scaled
like crazy. This playbook worked so well that Xiaomi went from phones to TVs, routers, scooters, washing machines, fridges, and basically every home
appliance you can imagine. And they eventually become a consumer hardware empire with a phone at the center of the spotlight. of like Apple, but they
were selling a robot vacuum. And then in 2021, they also decided to pull a crazy move, which is to
start making cars. More specifically, the smart EV business. With a plan to invest 10 billion dollars over 10 years, this is probably the biggest risk
they have ever taken. Because making home appliances are a bit more chill. First of all, they're not that big and second, they're not that expensive
compared to a car. But making cars, especially electric ones, you need new engineering, batteries,
motors, manufacturing lines, and basically a new supply chain problem that they need to solve. Which, coincidentally, solving the supply chain puzzles
is one of their strong suits. As you can see from their product range. But what's really impressive is that it only took them 3 years to start
shipping cars and in 2024, they already delivered over 136,000 cars, giving them a peak of 172 billion US
dollars market cap in 2025. So, So, efficiency kind of explains their insane acceleration when entering the AI side of things, or maybe not, cuz
clearly most things they have made or done are physical devices. LLMs, they are intangible. There's just no way. Well, today is April 22nd, 2026,
Xiaomi had just topped the open-source leaderboard with their latest model Mimo V2.5 Pro, beating labs like
DeepSeek, MiniMax, ZAI, Quinn, and even tied with Kimi. Like, how? Maybe when you have discipline like them and are insanely good at executing a
business goal, anything is possible. But, even the richest company Nvidia couldn't achieve this with all their compute. Same goes for Meta and many
other talented AI labs. So, how exactly did Xiaomi just somehow appear out of nowhere and pull off every
trillion-dollar company's dream in just 1 year? I wouldn't say there's like a key to success that they discovered, but this person called Lawful Lee
definitely made a huge impact on the development of their LLM research. With her publicly announcing and potentially hinting that she is leading the
core LLM team at Xiaomi in November 2025. Before she joined them, she was actually one of the key
people behind a lot of DeepSeek's research. From some of the earliest papers like DeepSeek LLM and MoE released in January 2024, and then DeepSeek
Coder V2, Coder V2, V3, and all the way to R1 that were released in January 2025, she was a contributing in all of them. Well, she seems to have left
DeepSeek before the release of V3, but she is still a contributor to both V3 and R1 as suggested on the
papers. So, the departure happened somewhere between July and December 2024. But again, the ability of being able to poach people doesn't naturally
mean you would make good models, as you can see from the case study of Meta's super intelligence lab. However, Xiaomi with her definitely took a
different turn than anyone would have expected. The AI initiative from Xiaomi called Mimo has a core LLM
team that so far only published five papers ranging from Mimo in May 2025, Mimo VL in June 2025, Mimo audio in December 2025, Mimo V2 Flash in January
2026, and High Spars, which is a novel attention paper in February 2026. Other than the original Mimo paper, Lufanli was the corresponding author for
all of them, which may imply that she is the one leading the LM side of research. What's even
cooler is that she even contributed to the embodied intelligence team that develops autonomous driving. But besides that, most of these model releases
have also been open weight releases. And the first one is Mimo 7B, which was trained from scratch as a reasoning-focused language model pre-trained on
25 trillion tokens with multi-token prediction and has a 32,000 context window using GQA and a
post-training recipe built around verifiable math and coding problems. For comparison, Qwen 3 AB that also came out around the same time as Mimo 7B
was pre-trained on 30 trillion tokens, and Mimo outperforms it on their official benchmarks. But for another comparison, DeepSeek V3 was only
pre-trained with 15 trillion tokens, with it being a 671-billion-parameter model that is 96 times larger than
Mimo and was able to reach near SOTA around the time it came out. So, while having a large token count for important, the quality factor of the token
still matters a lot. But it seems like a really solid model from the get-go, and them being able to have such a huge token count for pre-training in
their first release is also really impressive. The base model seems to be a decently popular model as
well, with over 127,000 downloads on Hugging Face. I mean, of course, it'll still not be Qwen 3 AB that has over 11 million downloads. But just a
month later, they have already expanded into vision language modeling. The new model called Mimo VL was built on top of Mimo 7B, adding a native
resolution VIT encoder and an MLP projector. Then it was trained across images, videos, OCR, GUI data, and
documents data. In the landscape at that time, where the only other VLM models were from Quinn and Google, Mimo instantly matched it and even
surpassed Quinn and Gemma and even become on par with GPT-4.0 on OCR tasks. And thanks to it being built on top of Mimo 7B, it retains strong language
performance that it seems to also beat GPT-4.0 across language-related benchmarks. The same goes for Mimo
audio, which is a solid model experiment published in December. The core idea is basically a scaled-out foundational audio model built around the same
next token prediction paradigm as LLMs. But unlike a normal LLM where the input is just text tokens, audio is much harder because you need to preserve
both the semantic content and the acoustic details. Like what you said, who said them, and the
speaker's tone. So, Xiaomi first built a 1.2 billion parameter audio tokenizer trained on 100 million hours of audio to convert speech into discrete
audio tokens while preserving as much information as possible. The tokenizer runs at 25 Hz, meaning it represents the audio every 40 milliseconds. But
each 40-millisecond slice is not just one token. It uses eight layers of residual vector
quantization, so each slice becomes eight discrete tokens. That gives you 25 slices per second times eight tokens per slice, or about 200 audio tokens
per second, which is a problem. Text is much more compact. If you transcribe 1 second of speech into text, you may only get a few text tokens. But if
you model the actual speech signal as audio tokens, that same second becomes around 200 tokens. So,
if you feed those directly into an LLM, the sequence length explodes very quickly. To solve this, Mimo audio adds a patch encoder and patch decoder
around the LLM. The patch encoder groups four consecutive audio slices together before sending them into the LLM. Since the tokenizer represents audio
25 times per second, and grouping every four slices reduces what the LLM sees down to 6.25 steps per
second. So, instead of forcing the LLM to process the full high-frequency audio token stream directly, it reasons over a more compressed patch-level
sequence. Then the patch decoder expands that lower-rate sequence back into the full 25 hertz audio token stream, auto aggressively reconstructing the
detailed audio tokens needed to generate the waveform, which means it can listen to audio and answer
questions. On top of that, for a normal text LM, few-shot prompting means you put a few input-output examples in the context and ask the model to
follow the pattern, right? Mimo audio here is trying to do the same thing, but with audio examples. So, for a voice conversion, you can give it
several paired examples showing one voice transformed into another voice, then give it a new utterance and ask
it to apply the same transformation. For speech denosing, you can give it noisy-clean audio pairs, then give it a new noisy recording and ask it to
output the clean version. For speech translation, you can give it paired English-Chinese speech examples, then give it a new English sentence and ask
it to produce the Chinese speech output. So, just like GPT-3 learned a new text task from a few
demonstrations all through skilling, Mimo audio is trying to learn a new speech task from a few audio demonstrations here, too, which is fascinating.
With this setup, they were able to get set of the art on speech intelligence and audio understanding benchmarks. While this idea may not be completely
new, this may still be the first open-source research targeting the few-shot capabilities of audio,
as most previous research like Kimi audio was only done up to 13 million hours of audio, which is nearly 10 times less than the Mimo audio and
probably the first one saw the emergent few-shot capabilities just like GPT-3 back in the days. But so far, most of this research has been building on
top of the Mimo 7B that they have built in mid-2025. And due to its size, it's still nowhere close to
compete with frontier models. That is not until the release of Mimo V2 flash in January 2026, which is a 309 billion parameter mixture of experts
model with 15 billion active parameters per token and trained on 27 trillion tokens and designed specifically for fast reasoning and agentic
capability. A completely different class of model compared to Mimo 7B that aims to compete at the frontier. It
now has 256,000 context window thanks to a different attention design that now relies on a 128 token sliding window attention and global attention
hybrid with a ratio of 5 to 1 utilizing group query attention. And what's really fascinating about this attention hybrid is that they actually picked
a 128 token sliding window instead of any other larger sizes like 512 tokens. So, in hindsight, we
would assume that the more context the better, right? But what they instead got was this weird result where after long context extension, the 128
token window actually performs better while the 512 token version degrades much more heavily. So, their interpretation of this result is that the
smaller window creates a cleaner division of labor where the sliding window attention layers are forced to
specialize in local transformations and the global attention layer specialize in long-range. As for the 512 token window, their guess is that the
sliding window attention layers would start to partially try to handle mid-range and long-range token relations themselves, so the roles becomes
blurrier which results in degraded performance. On top of that, hybrid sliding window attention was assumed
to be generally inferior in performance compared to global attention, but as shown in their ablation studies, it says otherwise pretty much catching
the whole field completely off guard with the difference as big as four points on average across long context benchmarks, too. They also included the
attention sync trick proposed by OpenAI in their GPT-OSS model and it also makes a four points
difference across the board. What the sync does is that the model basically gets an option to not strongly attend to any real token when the local
context is not useful. And other than its unique architecture experiments, they also applied newer methods to post-training where specifically I want
to talk about their multi-teacher on-policy distillation. I have already discussed this in my
DeepSeekV4 video, but the main idea is instead of forcing one model to learn everything from one mixed reward through RL, they train specialized
teachers and distill them. So, each domain teacher provides dense token-level guidance through KL rewards and this is also better than training RL
directly as the learning process is token-level instead of only final answer level. So, the model gets much
denser feedback than normal sparse RL. And from these techniques, you can tell that they not only have the good eyes on what's the frontier model
training methods, but also are keen at testing variations like sliding window attention that people are overlooking. They were also able to serve it
at a pretty decent price of 0.1 in and 0.3 dollar out per million tokens and hitting 150 tokens per
second on H200. This model is basically on par with DeepSeek V3.2 as of release sitting at only half of its size. But, of course, they are not going
to be able to beat DeepSeek's price as they are the real beast at infrastructure optimization. But, other than chasing new model releases, they also
aim to reduce the price publishing this rather experimental paper called HiSparse that answers the
question, what if the full attention layers can also tell the next layers which tokens are important? The idea is that these full attention layers
would look over the entire context so they can directly identify which token blocks are important. However, the full attention layer would act as the
high-quality selector because it already did the expensive full context lookup. Then, the following
sparse attention layers reuse those selected block indices instead of guessing important tokens themselves. They call this process Oracle token
selection. This is pretty new because most sparse attention methods only reduce compute while still keeping the full KV cache around. And if you throw
away the wrong tokens, you cannot recover them later. But, HiSparse tries to reduce both compute and
memory. The full layers picks the important blocks, the sparse layers reuses those choices, and they also reuse the KV cache from the full layer. And
they still keep a small sliding window branch for local context because sparse global retrieval alone is not enough. In their 80 billion parameter MoE
experiment, they use a very aggressive 1 to 11 ratio where only five out of 49 layers are full
attention. And yet, HiSparse still outperforms both full attention and hybrid sliding window attention baselines while reducing KV cache stores by
nearly 10 times. Even though this idea is cool, the models they released 2 months later, V2 Pro and V2 Omni, doesn't look like they incorporated this.
Compared to V2 Flash, Mimo V2 Pro was basically a bigger agentic model scaling past 1 trillion total
parameters with 42 billion active parameters and supports up to 1 million token context window. It also uses the same hybrid attention design, but
pushes the ratio even more aggressively from 5 to 1 to 7 to 1. However, we do not know how many tokens it was pre-trained on. And due to the same
attention setup for up to 256,000 context, the Mimo V2 Pro costs $1 input and $3 output per million tokens.
And for 256,000 to 1 million context, it costs double the amount. And unfortunately, it is not open weights. It is only available through their API.
It did, however, achieve rank two for the Chinese model when it just came out, sitting behind GLM 5 on artificial analysis. As for the Mimo V2 Omni,
it is built to process text, images, video, and audio together with dedicated image, video, and audio
encoders fused into a shared backbone. In one of their experiments, they put V2 Omni inside Open Claw, gave it a browser task, asked it to browse
right now for Xiaomi 17 advise, go to jd.com, compare sellers, negotiate with customer service, and complete an add-to-cart and checkout flow. A
really strong multimodal understanding model that is nearly on par with Gemini 3 Pro. It is also only
available through API, but seems to be limited to 256,000 context window. And to add on to that disappointment, a month later, they released a Mimo
V2.5 Pro. And this model is nuts. It basically topped the Chinese LM chart, including beating Qwen 3.6 Max, Deep Seek V4 Pro, GLM 5.1, MiniMax M2.7,
and tied with Kimi K2.6 with 54 points on artificial analysis. But it seems to suffer the same fate as
V2 Pro with it being closed sourced, even though they stated that it's open source in their announcement post. How could they? Well, as people were
disappointed with the lack of hugging face links, they dropped this rather goofy line and after 5 days, they actually uploaded it. This is not a tiny
model either with roughly the same model spec and architecture as MiMO V2 Pro, but now pushed it even
harder. While V2 Pro's hyper attention ratio is 7:1, V2.5 Pro actually reduced the ratio down to 6:1, but both still have the same tiny 128 token
sliding window. V2.5 Pro does have a much stronger training and post training recipe though with its main focus in the agentic and long horizon
coherence. I mean, not much more information was shared about how they did it, but I guess the benchmark would
usually reflect on that. On artificial analysis, it is now number one on long context reasoning for open weights and number three overall. Number one
on GTP valve for open weights which measures real-world economic value and sits at number five overall. Number one non-hallucination rate overall.
Number three on humanities last exam for open weights. Number two overall on instruction following.
Number one open-source model in frontier coding task on design arena and number three overall. And it is in the top five most efficient model. As for
its bad performing scores, it is pretty bad on omni science, so it seems like the pre-training could be a lot stronger and it lacks a lot of
scientific knowledge with its GPQA diamond score being pretty low. On Xiaomi's official benchmarks, you can
clearly see that it's just slightly behind big three, but beating every other model. Well, I mean Kimi K2.6 is not here, so it's not really every
other model. As the example for a long horizon task, they showed the model developing a full system-wide compiler in Rust from scratch. This is the
kind of project that would normally take a Peking University CS students several weeks, but V2.5 Pro
finished it in 4.3 hours across 672 tool calls and successfully passed through all 233 hidden tests. And the trajectory is really interesting, too.
They say the model first scaffolded the full compiler pipeline, then perfected the IR layer, then finished the RISC-V back end, then worked on
performance optimization. And when a refactor caused a regression around turn 512, it diagnosed the issue,
recovered, and kept going. Then, they show a second example where it built a full-featured video editor, with the final build containing 8,192 lines
of code, producing over 1,868 tool calls across 11.5 hours of autonomous work. But, on the other hand, the major killer for these long-horizon models
is that it is highly likely to cost a kidney if the model is not only expensive, but requires a lot
of thinking. However, Mimo is actually leading the token efficiency with respect to performance, using roughly 40 to 60% fewer tokens than Claude Opus
4.6, Gemini 3.1 Pro, and GPT 5.4 at comparable capability levels. And even though DeepSeek is cheap, there's still the fact that the token efficiency
of Mimo still costs so much less to run all of the artificial analysis benchmarks compared to
DeepSeek Pro Max and other models, being the best performer, yet the cheapest model out of them all. I mean, of course, I'm making this conclusion
based on benchmarks, but all my AI lord, I think the most jaw-dropping part about this release is that it is a true MIT license model, just like
DeepSeek. Even Kimi K 2.6 is modified MIT, which requires extra attribution for very large commercial
products. So, yeah, what a crazy way to break into the market and assert dominance. I am also insanely impressed by how they are able to achieve all
these within 1 year of their first model release. Who could have thought of all the Chinese AI labs, Xiaomi is actually rising the fastest. They
definitely have some sort of company culture or dedication that makes them stand out more than the others,
because if you want to talk about the budget, they are not the ones with the largest budget, even within China. And if you want to talk about shipping
fast, well, ByteDance, Kuai An, Kimi, and DeepSeek are also full of insanely talented researchers that are all working at an incredible speed. So,
this makes me really excited to see what the future of Xiaomi holds in this AI race. But, what do you
think? Let me know down in comments. And if you want to learn more about how LLMs work much more in depth, without being overwhelmed with math, you
should definitely check out my latest project intuitiveAI.academy, where it contains an intuitive explanation of all modern LLMs from the ground up,
including a lot of technical topics like distillation, which I just mentioned in this video. We cover
everything from the basics like the transformer architecture, all the way to more advanced topics like Laura, mixture of experts, and RLHF. And we
also just added a new advanced chapter on optimizers, which will bring you all the way from the classics to the current frontier techniques. So,
whether you're a student, software dev, founder, or just someone trying to pivot into AI,
intuitiveAI.academy gives you one clean place to build real technical intuition, and you can use the code summer for 25% off on a yearly plan. And
thank you guys for watching. A big shout out to Spam Maj, Chris Ladue, Deegan, Robert Zaviada, Marcelo Ferreira, Proof and Inu, DX Research Group,
Alex, Midwest Maker, and many others that support me through Patreon or YouTube. Follow me on Twitter if
you haven't, and I'll see you in the next one.