Google just revealed a model that
promises up to four times faster AI
responses on dedicated GPUs, [music]
then launched live voice translation
that can turn one language into another
while people are still talking. Xiaomi
jumped into the coding agent war with an
open-source tool that remembers huge
projects better than clawed code, and
OpenAI quietly moved closer to a
possible $1 trillion IPO. All right.
Now, Google just released diffusion
[music] Gemma, an experimental open
model that generates text differently
from normal AI models. Most chat bots
today write one tiny piece at a time.
They produce one token, then the next
token, then the next one. It is
basically like a typewriter. That is why
these models can feel like they are
typing the answer in front of you.
Diffusion Gemma does not work like that.
Instead of writing from left to right,
it starts with a messy block of
placeholder text [music] and improves
the whole thing over several rounds. It
is similar to how AI image generators
start with noise and slowly clean it up
into an image. But here, the noise
[music] is text. The model is cleaning
up words. That matters because a normal
AI model mostly commits [music] as it
goes. Once it writes a word, it usually
builds on top of that word. Diffusion
Gemma works on a 256 token canvas,
[music]
meaning it can look at a whole chunk of
text at once. So if something near the
end changes the meaning, the model can
still adjust earlier parts and correct
itself while the answer is forming. The
model itself is a 26 billion parameter
mixture of experts model built on
Google's Gemma 4 family, specifically
the 26BA 4B architecture. But it only
activates around 3.8 8 billion
parameters during inference. [music] So
it does not use the full model every
time. Google added a diffusion head from
its Gemini diffusion research and
released the model under Apache 2.0. The
speed is the headline. Google says
diffusion Gemma can generate text up to
four times faster on dedicated GPUs.
[music]
On a single Nvidia H100, it can go above
1,000 tokens per second. And on an RTX
5090, above 700. quantized, it [music]
can fit inside about 18 GB of VRAM. So,
high-end local machines are part of the
point. But this is not Google saying
this is our best writing model. Google
is clear that standard Gemma 4 is still
better when you need the highest quality
output. Diffusion Gemma is about speed,
local use, and interactive [music] work.
It is for things like inline editing,
fast drafting, code infilling, OCR,
document [music] parsing, agent
workflows, and structured tasks where
the answer depends on many parts fitting
together. A good example is Sudoku. A
normal leftto model struggles because
Sudoku is not just about the next
number. The whole grid matters. The base
diffusion Gemma model basically solved
0% of Sudoku puzzles, [music] but after
fine-tuning with Unsloth using a
supervised Jack's recipe, it reached
around 80% correctness. That shows why
the whole structure matters. There is
one more important detail. This speed up
makes the most sense for local and low
concurrency use. In a huge cloud system,
normal AI models can batch thousands of
users together and keep the hardware
busy. But when one person is running a
model locally, the GPU can sit underused
while the model slowly goes token by
token. Google is also making it
available through the usual developer
ecosystem. The weights are on hugging
face and it works with MLX, VLLM,
transformers, unsloth, Nvidia Nemo, and
Hackable Diffusion. Red Hat is
supporting VLLM, llama.cpp CPP support
is coming soon and Google worked with
Nvidia on Hopper, Blackwell, RTX Pro,
DGX Spark, DGX Station, and consumer
GPUs like the RTX 4090 and 5090. It also
supports NVFP4,
a [music] 4bit format that helps speed
things up with near lossless accuracy.
Now, Google's second announcement is
less strange, but probably more useful
for normal people. Google introduced
Gemini 3.5 Live Translate, a new audio
model for near realtime speech-to-Spech
translation in more than 70 languages.
The important part is that it does not
just translate after someone finishes
talking. It listens while the person is
speaking and generates translated speech
[music] only a few seconds behind.
>> Welcome to our tour of the beautiful
city of Cartahana.
>> Older translation systems often feel
like a stopandgo conversation. Someone
talks, everyone waits, the translation
comes, then the other person replies.
Gemini 3.5 Live Translate is meant to
[music] feel more fluid. It tries to
preserve the speaker's tone, pacing,
pitch, [music] and natural rhythm
instead of sounding like a flat robot
reading subtitles. The model
automatically detects languages so
[music] users do not have to manually
set every language combination. Google
also says it is built [music] for noisy
places like cars, airports, streets,
offices, [music] classrooms, meetings,
and video calls. The rollout is pretty
broad. Developers get public [music]
preview access through the Gemini Live
API and Google AI Studio. Selected
Google Workspace customers get private
preview access inside Google Meet this
month with a wider roll out later this
year. And regular users get it in the
Google Translate app on Android and iOS.
In the Translate app, people can connect
headphones and hear translated speech
across more than 70 languages. On
Android, Google is adding a listening
mode where you hold the phone to your
ear like a normal call and hear the
translation through the earpiece. Google
Meet is getting a major upgrade, too.
[music] The older speech translation
system supported only five languages and
mostly worked through English. The new
system supports more than 70 languages
and over 2,000 language combinations in
one meeting, so people can speak
different languages without everything
being forced through English first.
Partners are already testing it. Grab is
trying it for [music] communication
between drivers and travelers at
pickups. And Grab users make more than
10 million voice calls per month through
the app. Platforms like Agora, Fish Jam,
Livekit, Pipe Cat, and Vision agents are
also integrating the Gemini Live API. So
developers can build translation apps
without handling all the real-time audio
infrastructure from [music] scratch. And
because this is AI generated audio,
Google says the output is watermarked
with synth ID. People cannot hear that
watermark, but it is meant to help
identify AI generated audio later and
reduce fake audio abuse. Now, while
Google is working on speed and
translation, Xiaomi just made a serious
move into AI coding. Xiaomi's MIMO AI
team released MIMO codev0.1.0,
an open- source coding [music] assistant
that runs directly in the terminal. It
is available on GitHub under an MIT
license, installs with one command on
Mac OS and Linux, and can be installed
with npm on Windows. It also comes with
limited time free [music] access to MIMO
v2.5 through MIMO auto. The interesting
part is the problem it is trying to
[music] solve. Coding agents are often
good at short tasks, but they get worse
during long sessions. They forget
earlier decisions. They lose track of
the project. They stop remembering why a
file was changed or what the plan was.
Then the developer has to explain
everything again. Xiaomi's argument is
that bigger context windows are not
enough. Myo code is built around memory.
It has a persistent memory MD file for
project memory, session checkpoints,
scratch notes, and task progress logs.
Under the hood, it uses SQLite FTS 5
full text [music] search so it can
retrieve useful information across
sessions. There is also a separate
checkpoint writer sub agent. In simple
terms, the main coding agent keeps
[music] working while another agent
writes down the important decisions and
progress. So when the coding agent gets
lost, it can rebuild the situation from
structured checkpoints instead of
starting over. Myo code also has a
slashdream command that runs roughly
every 7 days to review old sessions,
remove duplicates, and [music] compress
useful information into long-term
memory. It has a distill function, too,
which can find repeated workflows and
turn them into automation. [music] Xiai
claims this helps most when tasks get
long. In its own benchmarks, Myo Code
with Myo V2 [music]
5 Pro scored 82% on Swebench verified,
[music]
compared with 79% for Claude Code with
Claude Sun at 4.6. On Swebench Pro, Xiai
claims 62% [music] versus 55%.
On Terminal Bench 2, it claims 73%
versus 69%. [music]
Xiai also says the harness itself adds
performance. When the same Mimo V2 5 Pro
model was used in both systems, MIMO
code scored 62% on SWEBench [music] Pro
compared with 57% in Claude Code and 73%
on Terminal Bench 2 compared with 68%.
The human test is even more interesting.
Xiaomi says it ran a double blind AB
test with 576 developers, 474 real
private repositories,
and 1,213
judged comparisons against Claude Code
using the same target model. Under 200
execution steps, the systems were about
tied. But after 200 steps, Mimo Code's
win rate went above 65%. Of course,
there are caveats. These are Xiaomi's
own numbers, not independently [music]
verified. MIMO code is not on the
official leaderboards yet. Xiaomi also
did not compare directly against OpenAI
Codeex CLI or Google Gemini CLI. On the
official Terminal Bench 2 leaderboard,
OpenAI Codeex CLI with GPT5, 5 is around
82.2%
and OpenAI claims 82.7%
well above Xiai's 73%. But on SUI Bench
Pro, OpenAI reports GPT55 at 58.6%
while Xiai claims 62%. Feature-wise, Myo
Code can read and write files, run
commands, manage [music] git and import
MCP servers, custom skills, and API
settings from Cloud Code. It has compose
mode where you press tab, describe a
bigger goal, and the system handles
planning, coding, testing, and review.
[music]
It also has voice control for loggedin
users powered by MIMO ASR and Tenvad so
developers can dictate instructions and
[music] say commands like send or
execute. The model behind it is
aggressive too. MIMO v 2.5 has 310
[music]
billion total parameters, 15 billion
active per inference and a 1 million
token context window. My v 2.5 Pro has
1.02 02 trillion total parameters
[music]
and 42 billion active parameters. Xiaomi
says the Pro model was trained with
harness awareness, meaning it was
designed to work well inside agent
systems with memory, tools, and [music]
context management. The pricing is also
aggressive. MIMO V 2.5 starts at 40
cents per million input tokens and $2
per million output tokens. MIMO V2.5 Pro
starts at $1 input and $3 output up to
[music] 256,000 context, then doubles
beyond that. Cash hits can bring input
costs down to 20 to 40 cents per
million. In the same pricing snapshot,
GPT 5.5 was listed at $5 input and $30
output, while Claude Fable 5 or Mythos 5
was [music] listed at $10 input and $50
output. My code also supports DeepSeek,
Kimmy from Moonshot, GLM from JIPU, and
any OpenAI compatible API. So, Xiai is
trying to build an ecosystem. This fits
the bigger pattern from Chinese AI
companies like Deepseek, Alibaba's Quen,
Miniax, Moonshot, and now Xiai, open
tools, capable models, permissive
licenses, and very low prices. And
finally, while all of this is happening
in products and models, OpenAI is moving
closer to the stock market. Reuters
reported that OpenAI confidentially
filed for a US IPO. That means the
company has started the process of going
public without revealing the full
details yet. OpenAI did not disclose the
offering size, terms, or final timeline.
It also said it may still take a while
because some things are easier to do as
a private company. But Reutder says
OpenAI has been targeting a valuation of
up to $1 trillion with a possible debut
as early as September. Anthropic also
said on June 1st that it had
confidentially filed for a US IPO after
reportedly raising $65 billion at a $965
billion valuation. SpaceX is also
pursuing public markets with reports of
a $75 billion offering at a $1.75
trillion valuation. OpenAI's numbers
show why investors are watching. Earlier
this year, it said it was raising $110
billion at an $840 billion valuation
from backers [music] including SoftBank,
Amazon, and Nvidia. It also said chat
GPT had more than 900 million weekly
active users and over 50 million
consumer subscribers. In March, OpenAI
said it was generating $2 billion in
monthly revenue compared with about $1
billion in quarterly revenue at the end
of 2024. But Reuters also reported that
OpenAI told investors it does not expect
to be profitable until 2030. There was
also the legal drama. OpenAI started as
a nonprofit in 2015, created a
for-profit arm in 2019, and later
announced plans to become a public
benefit corporation. Elon Musk sued,
accusing OpenAI of moving away from its
original mission. But in May, a US jury
ruled against him that removed a major
legal obstacle before a possible IPO.
OpenAI also renegotiated its Microsoft
partnership. Microsoft has invested $13
billion since 2019, but the new
arrangement reportedly gives OpenAI more
freedom to work with companies like
Amazon and Google. That's it for today.
If you enjoyed the video, make sure to
like, subscribe, and let me know what
you think in the comments. Thanks for
watching, and I'll catch you in the next
one.