Minimax M2.5 is the next iteration from

the previous M2.1 model and the broad

market implication of this model is

quite large. The model is a contemporary

to the state-of-the-art models like

Anthropic Opus 4.6, OpenAI GBT 5.2 and

Gemini 3 Pro. But what's really

impressive is the fact that each token

only activates 10 billion parameters out

of the 230 billion parameters in total.

Now, this kind of sparsity isn't

anything new, per se, where the model

only activates around 4% of its brain

power to generate output. But when you

compare Miniax M2.5 against other models

like GLM5, Kimik 2.5, and Deepseek V3.2

Special, you start to see why this is

such a big achievement. Now, you might

wonder why this even matters. And here's

why. In order to fit trillions of tokens

of training data into the model, you

need to pick the right size of the

model. And Miniax chose 230 billion

parameters in size for their M2 series.

And now once this knowledge is packed

into the model's weights, you also need

to retrieve that information out of the

models efficiently during inference. And

in the case of Miniax M2 series, they

chose to activate only 4% of the entire

parameters at 10 billion parameters.

Meaning every token that the model

outputs only uses 4% of the model's

entire weights through mixture of

experts. Now, even at the 10 billion

parameter sparity, they were able to

achieve 80% in the Sweetbench verified

benchmark, which is neck-to-neck with

Enthropic's most recent model, Opus 4.6.

And because the memory footprint is so

low at 10 billion parameters, MiniAX is

not only able to serve the model at 3%

of the cost of Opus 4.6 in output

tokens, but they also host the model

nearly twice the speed at 100 tokens per

second. Now looking at the progression

going from M1 to M2, M2.1 to M2.5 at

each progression, they were able to up

the performance given their release

cycle of around 50 days during each

release. And they were able to do this

while maintaining the cost at 30 per

million input tokens and output token

cost at $1.20 per million output tokens.

Here's another way to look at why this

is such a big deal. Ever since Moldbook

and Open Claw became viral, one

impression that left on people's minds

was this idea called always on agent,

where you can have an agent that's

always listening, always doing some

background tasks, and always ready to

work on things. But one of the problems

with this was when it comes to cost. As

much as having an always on agent was

extremely cool, the total cost of

ownership was just too high for a lot of

people. So, assuming you're running an

agent 24/7 for an entire year that's

constantly outputting tokens, just like

leaving the tap on in the bathroom, the

cost will look different depending on

what model you choose. And as you can

see, similar to hiring an employee for a

job, you have different models at

varying wages in terms of salary per

year, but also varying levels of

intelligence, which in this case, the

intelligence is sampled by SWEBench

benchmark. And for Miniax M2.5, you're

only paying $1,892

per year in salary, assuming they run at

50 tokens per second for the entire

duration. And even at that cost, the

model is just as capable as the OPUS 4.6

in performance, but at a 2 to 3% cost of

intelligence. So what is a broad market

impact for something like this? One of

the biggest debates when it comes to AI

bubble was the idea around the falling

cost of intelligence. Meaning as the

cost to run AI models gets cheaper and

cheaper, the demand for graphics cards

will decrease because AI models just

like the Miniax M2.5 are becoming more

and more efficient, which means we can

more or less do more with less graphics

cards. So when people saw that companies

like OpenAI, Oracle, and XAI are

spending tens and billions of dollars

into graphics cards, while the cost of

intelligence is dropping, it caused

panic in the market of potentially

overprovisioning graphics cards as

models gets cheaper and more efficient.

And the argument against this was a

Jevans paradox where as resources get

cheaper, the demand actually grows as a

consequence. In other words, the cost of

intelligence dropping is actually a good

thing because now more people will want

to use it. And Miniax M2.5 is really one

of the first signals towards this

paradox where the cost of intelligence

is actually nearing the cost of

electricity at this point which means

the aggregate demand for AI and agents

should also increase. Now, I think we

all envision someday that models like

Miniax M2.5 will someday be able to run

on our phone. And well, we're just not

there yet. In fact, running that Miniax

M2.5 still requires up to 400 GB of VRAM

to fit both the model's weights and the

KV cache for inference. But just like

how quantized version could run locally

for the previous M2.1 model which share

the same architecture. Since Miniax

M2.5's models weights will be open

source shortly, we'll soon be able to

run this locally at home at a much lower

precision. You can try the Minax M2.5

model on the Miniax coding plan. Check

out the link in the description for 12%

discount for both new and existing

users.