---
格式版本: 2
标题: "How We Scaled Kimi K2.5 S81695 | GTC San Jose 2026"
原文链接: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/"
发布日期: "2026-08-11"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "Published Time: Tue, 11 Aug 2026 16:37:09 GMT"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 4
发布时间严格候选数量: 1
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-12T16:09:55+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-12T16:04:00+08:00"
入库时间: "2026-08-12T08:09:55.379Z"
来源平台: "固定入口"
搜索渠道: "fixed_url"
搜索词: "https://www.nvidia.com/gtc/"
匹配关键词:
- "GPU"
- "NVLink"
- "performance"
- "throughput"
- "AI"
相关厂家:
- "NVIDIA"
相关专家:
[]
内容类型: "网页"
抓取工具: "Jina Reader"
清洗工具: "Jina Reader Markdown + Defuddle/Readability 正文提取"
原始附件:
[]
AI优质: "否"
AI打分: 25
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "演讲内容为Kimi模型算法优化与Agent架构，未涉及超节点/AI Rack/机柜级硬件、供电、散热、互连等主题，与项目关注范围无关。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-12T16:10:05+08:00"
AI主题相关性: 0
AI来源权威性: 12
AI新颖性: 8
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 5
AI摘要: "在GTC San Jose 2026上，演讲者分享了扩展Kimi K2.5模型的方法，重点通过优化token效率、上下文长度和智能体群三个维度提升性能。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:40:51.452Z"
采集批次: "2026年8月12日13点00分43秒"
采集批次ID: "20260812-130043-786"
去重键: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695"
---
Title: How We Scaled Kimi K2.5 S81695 | GTC San Jose 2026
URL Source: https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/
Published Time: Tue, 11 Aug 2026 16:37:09 GMT
Markdown Content:
Visit your regional NVIDIA website for local content, pricing, and where to buy partners specific to your country.
[Continue](https://www.nvidia.com/)
[**GTC Berlin** October 20–22](https://www.nvidia.com/en-eu/gtc/) | [**GTC 2027** March 15–18](https://www.nvidia.com/gtc/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [](https://www.nvidia.com/en-us/account/)
* [Log In](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)[LogOut](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
PLATFORMS
other links
[](https://www.nvidia.com/gtc/)
Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
[Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/# "Menu")
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/# "Menu")
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81695/#)
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
[Video 1](blob:https://www.nvidia.com/87536398-1903-4705-be0e-ca7d260e8732)
Loading
Play
00:00
Play
Seek 10 seconds backwards
Seek 10 seconds forward
00:00 / 00:00
Mute
Press TAB to open volume control Use the arrows to control the volume
Enable Captions
Settings
Play on TV
Turn on Picture in picture
Show Full screen
Transcript Powered by AI
X
X
00:12
Hi everyone, thank you so much for the introduction,
00:15
it's great to be here, it's great to have this opportunity
00:20
to share with you guys some of our latest progress and efforts.
00:26
So one of our major pursuits is to build better open models.
00:34
And we believe in democratizing intelligence.
00:37
With open models, you can deploy anywhere.
00:39
It can be on your local servers.
00:41
It can be on the cloud.
00:42
And you can access every single bit of the weights in the model
00:47
instead of just using a black box.
00:50
And this is one of the slides that I took from Jensen's
00:54
talk earlier this year at CES.
00:56
So as you can see, open models are quickly closing the gap
01:00
with proprietary models and it's reaching the frontier.
01:05
And we believe that with better met open models, we're going
01:07
to make intelligence more accessible to anybody in the world,
01:12
in every corner of the world.
01:15
But open models cannot be just open.
01:18
They have also to be great.
01:22
So in this talk, we're going to discuss how we
01:25
make open models great.
01:28
So as we know, scaling is a primary driver for a lot
01:34
of progress, maybe all of the major AI developments that we have
01:38
witnessed in the last few years.
01:41
And here, we're going to discuss how we scale our model in
01:44
different dimensions.
01:46
So on the left-hand side, the first figure you see here is
01:49
kind of the standard scaling law.
01:52
So on the x-axis, you have the log of the number of training
01:56
tokens, and on the y-axis, you have the log loss.
02:00
And as you scale the number of training tokens, you
02:02
get a lower loss.
02:03
But here the point is, we're not going to just scale the
02:07
number of training tokens, but we also want to improve the token
02:12
efficiency, meaning that we want to move this curve to the left-hand
02:17
side so that we can achieve a lower loss, a much lower loss, using
02:22
the same number of training tokens.
02:25
And this can be achieved by having better architectures
02:28
and optimizers, as we'll discuss in our later slides.
02:33
And the second scaling dimension that we're very interested
02:36
in is to scale the context length.
02:39
So as you can see in the second figure, if we increase
02:43
the context length,
02:45
Then we can have a much higher accuracy in terms of predicting
02:50
the token loss at a given position.
02:55
And this means that we can increase the capability of the
02:58
model to achieve more complex tasks by increasing the context stance.
03:04
So this is the second scaling dimensions that
03:06
we're going to talk about.
03:08
And the third scaling dimension is the number of agents.
03:11
So we introduced this new learning paradigm of agent swarms, where
03:15
we don't just rely on a single agent, but we also orchestrate a
03:21
swarm of agents that can accomplish the subtask in parallel, so that
03:25
we can increase the task capacity.
03:28
And we can translate all of this into the language of agents.
03:32
So if you look at token efficiency, it's mostly about having a
03:36
stronger prior so that you can have more efficiency when you do agent
03:41
RL to search for a better solution.
03:45
And when you think about long context, it's mostly about
03:48
increasing the context so that you can have a longer-running agent.
03:51
It can probably run for days or even weeks or months to accomplish
03:57
more tasks, more complex tasks.
04:01
And for agent swarms, it's another dimension that's added,
04:05
and at the end of the day, we're going to have a swarm of agents.
04:08
That each of them have a super long context, and each of them have a
04:11
very strong prior for us to search in this entire agent RL system.
04:19
All right, so we're going to start from token efficiency.
04:23
So this is one of the most classical figures in the history
04:28
of machine learning.
04:29
So it's taken from Kaplan et al.
04:32
And it basically says that if we scale proportionately
04:35
the number of training tokens, the model parameters, and
04:39
also the amount of compute, we can get lower and lower loss.
04:43
And this is one of the major breakthroughs that
04:47
the entire community.
04:48
has achieved in the last few years to get better intelligence.
04:53
But here, what we're interested in is to have better and better
04:58
token efficiency.
05:00
And here's the thing.
05:02
So one thing that I would like to emphasize is that token efficiency
05:07
is not just about efficiency.
05:09
It's actually also about improving the upper bound of intelligence.
05:14
So here's why.
05:16
So suppose you have, say, 50 trillion tokens, 50 trillion
05:22
high-quality tokens, and then you apply this new optimizer,
05:26
maybe the new optimizer, and then all of a sudden, you
05:29
have a two-time token efficiency.
05:31
So it means that
05:33
It's almost like magic that you get equivalently 100 trillion
05:37
tokens and nowadays we are scaling towards the data war
05:42
and we're hitting the data war and the amount of high quality data
05:47
is quite limited and if we suppose that is a constant amount then
05:51
we increase the token efficiency
05:53
It means that we're going to get better intelligence out of it.
05:57
It's not just about infrastructure efficiency.
05:59
It's about better intelligence.
06:03
So this is why we spend a lot of efforts in this aspect,
06:08
because it's going to push the frontier of intelligence.
06:12
And Mule Optimizer is one of the things that we have heavily
06:15
invested in since last year.
06:19
So it's a second-order optimizer.
06:22
And basically, every single gradient update is transformed
06:27
in a way that each entry is orthogonal to each other.
06:31
And this is very different from the traditional Adam optimizer.
06:35
And if you implement this optimizer properly, you can get a two-time
06:39
token efficiency improvement.
06:42
So we are the first work, we published the first work,
06:48
to demonstrate that our own optimizer is actually scalable
06:52
for LLM training.
06:53
And these are two key techniques that we employ to make it
06:58
effective for large-scale training.
07:00
So one of them is weight decay.
07:02
It is critical for scaling to larger models.
07:05
And the second is we want to ensure a consistent RMS
07:09
update compared to Atom.
07:10
So we have this adjustable coefficient that is applied to each
07:16
update, so that the resulting RMS is going to be comparable to Atom.
07:22
And to make MIOM memory efficient across all these NVIDIA GPU
07:28
clusters, we also developed a distributed MIOM optimizer
07:32
implementation that partitions the states across the data
07:36
parallel group so that we can have a very efficient implementation
07:41
for the MIOM optimizer.
07:43
And these are some of the results that were presented in the paper.
07:47
So as you can see, with the same number of parameters
07:51
and the same number of training tokens, we just replaced the
07:54
original add-on W optimizer with the new Mule optimizer.
07:58
It's going to improve the performance across
08:01
the board significantly.
08:04
But there was this new challenge that we encountered when we
08:09
tried to scale data further, when we tried to scale muon
08:13
for a 1 trillion parameter model, we encountered a new
08:16
issue about training instability.
08:19
So, as you can see on the left figure...
08:22
We observed that the MaxLogix quickly explodes
08:27
and quickly exceeds 1,000.
08:29
And the typical values for training for this MaxLogix is about,
08:36
say, 50 or maybe less than 100.
08:39
But for Mule, it quickly exceeds 1,000.
08:43
And at the same time, we observe training divergence
08:47
on the left-hand side.
08:47
If you look at the training loss, it goes down a bit, but then
08:51
at the end of the day, it explodes, and it cannot converge as expected.
08:56
So this is one of the technical challenges that we have to adjust.
09:01
And the solution to this is to introduce this new
09:04
technique called QKClip.
09:06
So basically what it says is that for each attention head in
09:10
this entire neural network, we're going to, in the forward pass,
09:13
we're going to compute the MaxLogit and then we're going to calculate a
09:18
dividing factor that can be applied to each key projection, as well
09:23
as the query projection, so that we can sort of clip the maximum value
09:30
of the query and the key to sort of Constrainted into a given range.
09:36
So that we're not going to have explosion anymore.
09:40
So these are some of the empirical results.
09:42
On the left-hand side, there are two curves, but they are strictly
09:46
overlapped with each other.
09:47
So these are the training curves before and after applying
09:51
the clipping technique.
09:52
So you can see the clipping technique does not affect
09:56
The training loss decrease at all, but on the right-hand side, if we
10:02
inspect the intermediate metric, if we inspect the MaxLogit, it's going
10:07
to be effectively constrained.
10:10
So it first explodes, as before, but at the value of 100, it's
10:16
going to be clipped at a constant value for a long time, and then
10:19
after a certain number of steps, it will just naturally go down.
10:23
So, the neural networks sort of find a way to constrain the
10:28
maximum value of the MaxLogic to ensure a stable training process.
10:34
And at the same time, it doesn't affect the training convergence
10:38
as shown in the left figure.
10:41
So we employed this technique in our K2 model training and
10:46
successfully scaled it to 1 trillion parameters.
10:50
And this is the first example of large-scale muon training
10:55
in the history of machine learning.
10:58
And the second dimension that we're very interested in is long context.
11:04
So this is another figure.
11:06
It's probably less known.
11:08
It's one of the hidden gems in these papers.
11:12
So instead of just pushing down the training laws by training
11:16
on more tokens, it has some insights from another perspective.
11:21
So as we can see, this is a comparison between
11:24
transformers and LSTMs.
11:27
So on the left-hand side, we can see that transformers
11:29
achieve a lower training loss given the same number of parameters
11:33
and the same number of training tokens as expected.
11:36
And this is why transformers become the de facto architecture
11:41
that people are using right now.
11:42
But on the right-hand side, it's really interesting to
11:45
see that transformers are actually better because it can
11:49
improve through the whole context.
11:52
So the x-axis is the token index in context.
11:54
And if you increase the token index, you can see that the
11:57
training loss of transformers actually dropped by a lot.
12:02
If you just continually increase the context length, the loss
12:06
just continuously drops down.
12:08
But if you look at the curve of LSTM, it just is saturated
12:13
after a certain number of tokens.
12:16
It means that transformers have this better capability
12:19
of capturing longer context, and this is what makes it
12:24
better, because if you go back to like 10 years ago, people used LSTM
12:30
for tasks like machine translation.
12:32
But it is not good for, for example, understanding entire code
12:35
base or running a super long agent trajectory to solve, for example,
12:42
writing Linux kernels from scratch.
12:45
It's not going to be accomplished by LSTMs.
12:47
So this is a very much needed capability in the era of agents,
12:53
because tasks are becoming harder and harder, and we
12:56
need longer and longer contexts.
12:58
So the research idea here is to develop a better architecture
13:03
so that we can efficiently scale to a longer context length and at
13:09
the same time achieve a lower per token loss at larger token indices.
13:16
And this is the motivation for which we introduced this new
13:21
architecture called Kimi linear.
13:24
And it contains this new linear attention variant called Kimi
13:31
delta attention, which improves the original gated delta rule,
13:37
GDR, by improved recurrent memory.
13:40
I will show the details later.
13:41
And at the same time, we're going to mix linear attention
13:44
layers with full attention layers using a 1-2-3 ratio
13:48
so that you can balance between this long context capabilities
13:53
and at the same time having a more efficient implementation.
13:59
So this is some of the formulation.
14:02
The basic idea is simple.
14:03
If you look at linear attention, in the original formulation,
14:08
the memory is going to be global.
14:09
So there is a global single decay factor that
14:14
is applied along the way.
14:16
So it means that
14:18
Basically, there are only two cases.
14:22
In one case, you're going to forget basically everything,
14:25
and you're not going to retain any information.
14:27
In the second case, you can choose to retain almost everything,
14:31
but at the same time, you don't have the capability to
14:34
leave out some of the unnecessary information in this long context.
14:38
So we introduced this key idea of having a fine-grained
14:43
decay factor as shown in this highlighted alpha term.
14:47
So instead of being a scalar, it's going to be a diagonal
14:52
matrix which controls the decay rate for each channel so
14:57
that we can have two possibilities.
14:59
For some of the channels, we can decay really, really
15:02
slow, meaning that we can retain this long-context information
15:06
across a very long range.
15:09
And at the same time, for the other channels, we can
15:12
quickly forget the information from the past indices to refresh
15:17
it and observe new information.
15:20
And this is to increase the expressivity of this model.
15:27
And of course, to leverage modern GPUs, we have to use
15:31
this chunkwise formulations so that we can parallelize
15:36
the computation of modern GPUs.
15:38
So the first equation here is the chunkwise
15:42
formulation of Kimi linear.
15:45
But as you can see, this is going to bring massive infrastructure
15:50
challenges because of this newly introduced alpha term.
15:54
Because now it is a matrix instead of a scalar, it cannot
15:58
easily be factored out.
15:59
So to achieve an efficient implementation, we rewrite
16:04
the entire equation into the bottom three equations.
16:09
So we introduced this matrix inversion operation, as
16:13
well as introducing
16:15
The cumulative decay factor, so that we can implement this
16:19
entire thing in parallel, without sacrificing any efficiency.
16:25
And more importantly, this is not an approximation, it's
16:29
an exact mathematically equivalent formulation, so that we can achieve
16:34
much efficient implementation without sacrificing any loss
16:38
in terms of performance.
16:41
So it's going to be as efficient as previous linear attention
16:47
variants, but at the same time, much more expressive.
16:51
So these are some of the results that we obtained
16:53
using a fair comparison.
16:56
So on the left-hand side, we see the performance on
16:59
two different types of tasks.
17:01
So MMMU is a short context task.
17:04
So for short context tasks, Kimi linear achieved
17:07
a better performance compared to MLA and GDN.
17:12
And at the same time, for longer context tasks such as RULER,
17:16
Kimi linear is also better than the other variants while being much
17:23
more efficient compared to MLA.
17:26
And when we scale the context further to, for example, 1
17:29
million tokens or even longer, it's going to be much more efficient
17:34
compared to the baselines.
17:36
And this is also the first architecture that can outperform
17:41
full attention across the board, including short context tasks, long
17:46
input tasks, and long output tasks.
17:50
So these are two key dimensions that we are interested in.
17:55
And the third dimension is the agent swarms.
17:59
So here is a diagram to showcase how we designed this agent
18:04
swarm paradigm to solve some of the more complex tasks
18:09
compared to single agent paradigms.
18:12
So here we have an orchestrator.
18:15
Or you can call it a main agent.
18:17
It's responsible for orchestrating tasks.
18:20
It has different options.
18:22
For example, we can spawn a group of sub-agents and assign
18:25
new tasks to these sub-agents.
18:28
Or you can collect the results from the return of these sub-agents.
18:33
And you can sort of perform this process in an iterative way.
18:38
And at the end of the day, you can accomplish a more complex task
18:43
compared to using one single agent.
18:45
And it's analogous to human society.
18:48
For example, if we build a company,
18:50
We need different roles and we need, for example, an orchestrator
18:55
or maybe we need a CEO to compose and assign the tasks
18:59
to different roles.
19:00
And at the end of the day, the entire organization
19:03
is going to have to move towards this same goal.
19:07
And here, for example, in this case, we have, maybe
19:10
you have the AI researchers, you have the web developers,
19:14
you have physical researchers, and they can study different topics.
19:17
And at the end of the day, you just collect the results
19:20
and spawn a group of fact-checkers and web developers and file
19:24
downloaders to assemble the results into a single report.
19:32
And this is another perspective to look at this new paradigm.
19:38
So the x-axis is the complexity of the task.
19:44
And the y-axis is the execution time.
19:46
And the complexity of the task is measured by the accuracy
19:51
of a group of models on such tasks.
19:55
So we can see with Agent Swarms, it's going to substantially
20:01
reduce the execution time compared to single agents.
20:08
It's going to be more effective, and this means
20:11
that we can scale this Agent Swarm paradigm Time to...
20:15
For example, if you run Agent Swarms with 100 or maybe even 1,000
20:20
sub-agents, you can accomplish a complex task within a certain
20:25
period of time that is tolerable for producing real economic value.
20:33
And we can certainly scale it in different dimensions.
20:36
We can scale the inputs.
20:38
For example, we can download and read hundreds of sources
20:42
or even maybe thousands of doses in parallel.
20:45
Or we can output, write a 100-page literature review in parallel.
20:52
Or we can take actions at scale.
20:54
We can perform data analysis for 10 different tasks.
20:58
And also it is orchestration at scale.
21:01
You have to learn to design subtasks and aggregate the results.
21:06
And technically, we define some new objective functions
21:10
to guide the learning process of our agent swarm system.
21:15
So there are three reward functions, reward objectives
21:20
that are considered here, compared to the conventional
21:25
single agent IRL learning.
21:27
So the first term is what we call the instantiation reward.
21:32
It incentivizes sub-agent instantiation to prevent
21:37
this serial collapse phenomenon from happening.
21:42
So basically, we don't want it to default to single-agent execution.
21:47
We want to encourage the parallel executions, especially when we...
21:54
When it's early stage in training and of course we can decay
21:59
the weight for this instantiation reward term over training course
22:04
because when it learns parallel execution we can reduce the weight.
22:10
And the second term here is Finish Reward.
22:14
And it is used because we observe one of the things in training
22:19
that some of these subtasks are just created but never finished.
22:25
So it's almost like it's going to hack the first term.
22:28
By just spawning a bunch of sub-agents, and the task might
22:31
be too complex, or maybe the task just doesn't make sense.
22:35
And here, we use this Finish Reward to basically encourage
22:39
that each of the sub-tasks should have a relatively high
22:44
ratio of completion.
22:46
Instead of just spawning a bunch of pseudo-tasks, we
22:50
need it to be meaningful.
22:52
So this is the second term that we use and of course we use
22:55
the same, you know, decay strategy.
22:58
We use the relative high weight at the beginning of training
23:01
and we decay it to a relatively low weight at the end of training.
23:05
And of course the third term is the standard term.
23:07
It's the outcome reward.
23:09
It's going to measure whether the entire task is completed and
23:15
then we're going to add these three terms in our reinforcement learning
23:21
And of course, we have to build the entire infrastructure
23:24
because right now you need to support the parallel execution
23:28
and they will need to support different reward functions
23:31
and to maximize the efficiency of the entire agent swarm IO system.
23:38
So here are three different things that we have tried scaling.
23:45
The Muon Clip Optimizer improves token efficiency.
23:49
And Kimi Delta attention in the Kimi linear architecture
23:53
improves long contacts.
23:55
And we also have the agent swarms paradigm to further
24:00
create a new dimension of scaling.
24:03
And all of this put together, we created Kimi K2.5, a new
24:08
model that we just released over one month ago.
24:12
Here's a short video to demonstrate some of its capabilities.
25:19
So yeah, there are a lot of interesting things, capabilities
25:23
that we discover from the model.
25:25
For example, it merges the visual capabilities with
25:29
coding capabilities.
25:30
So a lot of new things just emerge out of it.
25:34
It can read a video and then produce a website that sort
25:39
of replicates or style transfer The original video.
25:43
And all of this are due to successful and stable training
25:49
at the pre-training stage.
25:50
So this is also one of the most beautiful curves that
25:55
I observed in my life.
25:56
So this is the training curve of the K2.5 based model.
26:02
So as you can see, it went through over 15 trillion tokens and
26:05
of course in K2.5 we additionally trained another 15 trillion
26:09
tokens and the entire training process is just so stable,
26:14
there's no lost spike, especially when we introduced this new
26:18
Muon optimizer, we didn't observe any spike and this
26:21
smooth, stable training process produces a very stable outcome.
26:27
A very strong-based model that we can fine-tune on top of
26:30
it to achieve new capabilities as we introduced and saw in the video.
26:37
And this is also, of course, a trend on NVIDIA H800 GPUs, and you
26:42
should know in this H800 cluster contains two TB RAM and 8 GPUs.
26:50
They are connected by NVLink.
26:52
And another key innovation of KeyMeK 2.5 is that it is
26:59
the first open model with native joint vision text capabilities.
27:04
So if you look at previous open models, usually their
27:08
visual capabilities are added on top of a text base, meaning
27:12
that, for example, if you train the text models for
27:14
20 trillion tokens, and then on top of it, you do another 2 trillion.
27:19
Sort of a post-training process to add additional
27:23
visual capabilities on top of it.
27:25
But for K2.5, it's different in the sense that we fuse
27:30
the training process of vision and text from day one.
27:34
So it's called early fusion here.
27:36
We start from 0% of the progress.
27:39
So from day one, we're going to merge the vision and text
27:43
tokens and as shown in our preliminary experiments, it
27:47
outperforms our late fusion.
27:50
And some of the new capabilities that we observe also come
27:53
from this training recipe.
27:55
For example, if you want to do...
27:57
Vision to kill, you really have to merge vision and text into
28:02
a single brand to achieve that.
28:04
If you separate these two brands, it's not going to happen.
28:07
You have to align these two modalities into a shared embedding
28:11
space, a shared representation space so as to achieve this.
28:18
And another interesting thing that we observe is that these
28:22
two modalities can actually enhance each other.
28:25
So that's been long been a challenge that if you add
28:30
vision capabilities into a text model, it's going to
28:33
somewhat hurt the text performance.
28:37
But here we found that if you train it properly, these two modalities
28:41
can actually enhance each other.
28:43
So this is one of the key findings that
28:45
We observe in our training, so first, vision improves text.
28:50
So this is so interesting, so before Vision RL,
28:54
The performance in the first column, and then we have the
28:57
performance after Vision RL.
28:59
So here, Vision RL refers to a process that we only
29:04
use vision tasks.
29:05
So there is no task involved here.
29:09
We only have vision tasks.
29:10
For example, we teach the model how to count, how to answer
29:14
some of these visual QA problems without any, for example, math and
29:20
encoding problems in this space.
29:22
But we observe that it's going to improve the performance
29:25
for even, you know, reasonably heavy text tasks.
29:29
And on the other hand, text also improves vision.
29:32
If you have a very strong text space, you actually don't
29:35
need any vision SFT data in the training process.
29:39
And this is the approach that we adopt.
29:41
So it's called Zero Vision SFT.
29:44
Basically, we don't have...
29:45
We have basically zero vision SFT data, and the only SFT data
29:49
that we have is the text SFT data.
29:52
And then we do a joint RL over text and vision, and
29:56
you can see that we can achieve almost state-of-the-art performance
29:59
across the board on vision tasks without any vision data.
30:03
So it's clear that if you have a strong text space, it's also
30:07
going to improve the vision if you align these two modalities into a
30:13
shared space in your pre-training.
30:17
And also, these are some of the examples of
30:26
As I've shown in the video, it demonstrates strong capabilities
30:30
of visual design and front-end coding, and this also emerges from
30:35
our vision text through training.
30:39
So after all this, so this is all about Kimi K2.5.
30:44
And as you probably know, we released our new architecture
30:50
yesterday in our tech report.
30:52
It's called Attention Residue.
30:54
So here I'm also going to briefly talk about our new
30:58
work, which serves as a sneak peek into our next generation
31:02
architecture that we're probably going to adopt in our later models.
31:07
So here the motivation is quite simple.
31:11
Can we apply some of our techniques that we use in the temporal
31:16
dimension, and we just take some of the inspirations and then
31:21
apply it to the depth dimension?
31:26
And it starts from this residual connection.
31:28
So I still remember listening to Kimi's talk.
31:34
at a tutorial in ICML 2016, 10 years ago.
31:39
So it was a brilliant idea.
31:40
So basically, before ResNet, nobody was able to train deep networks.
31:47
If you increase the depth, if you increase the number
31:49
of layers for neural networks,
31:52
Nobody was able to train it because you observe this gradient
31:55
exposure and gradient vanishing, all these stability issues.
31:59
But then after the introduction of ResNet, we can train an
32:04
arbitrarily large number of layers.
32:06
You can stack as many layers as you want, and you don't
32:10
have to worry about the training stability issue and stuff.
32:14
And as discussed in Ilya's talk two years ago,
32:18
It basically says that residual connection is a variant of
32:23
LSTM, but just rotated 90 degrees.
32:27
So how do you understand this?
32:28
If you look at LSTM, it's a variant of recurrent net, right?
32:33
And it's a recurrent model process.
32:36
So we're going to take the hidden states from the last step,
32:41
and then we're going to have some gating mechanism, some function
32:45
to produce the current states.
32:48
And if you look at the depth dimension, the register connection
32:52
is basically the same.
32:53
We're going to take the outputs from the last layer, and then
32:56
we're going to apply some sort of function on top of
32:59
it to produce the current outputs of the current layer.
33:06
It's just the formulation is different.
33:08
For example, for a residual connection, we're going to
33:10
use a fixed addition.
33:11
We're going to add the previous hidden state
33:18
with the current output.
33:20
It's just the formulation that's different.
33:22
But the basic idea is the same.
33:23
It's a recurrent that applies in the dimension of death.
33:28
But on the other hand...
33:30
We can think about reformulating this function, instead of
33:36
having an LSTM, can we have an attention in the dimension
33:40
of that, and it's going to create new possibilities because attention
33:45
has been demonstrated to be so successful in the transformer era.
33:50
So what we're going to do is not just to take the last
33:53
hidden state, but we're going to consider all the previous
33:56
hidden states and use the attention operation, the attention mechanism,
34:00
to assemble and aggregate all of these previous hidden states
34:05
to compute the current state.
34:07
So this is exactly attention rotated by 90 degrees.
34:13
We view it as a natural generalization of residual
34:17
connections in the LSTM analogy.
34:21
Okay, and here is the detailed formulation.
34:24
So, on the left-hand side is a standard residue.
34:27
As I said, it is basically LSTM rotated by 90 degrees.
34:31
And the second figure is attention rotated by 90 degrees.
34:35
So, what we did is to collect all the previous hidden states
34:39
and have a simple attention operation on top of it to
34:43
produce the current layers outcome.
34:45
And of course, to increase the efficiency, to reduce
34:49
the infrastructure, for example, communication and memory overhead,
34:53
we also designed a new variant called block attention residue
34:57
on the right-hand side.
34:59
So basically, the idea is also simple.
35:02
We're going to divide all the layers in the neural networks
35:06
into multiple blocks.
35:08
For example, each block can contain, say, 16 layers, or
35:11
it can contain maybe four layers.
35:13
And then for each block, we're going to apply this
35:18
attention residue only on the output of each block.
35:21
But within each block, we also still adopt this standard residue.
35:25
So this is going to reduce a lot of overhead while having minimal
35:30
loss in terms of training accuracy.
35:33
And these are some of the impressive results that we
35:35
achieved on this new architecture.
35:39
So, on the scaling law, we can improve the token efficiency
35:44
by 24%, meaning that if you have 50 trillion high-quality
35:50
tokens, now you just magically have over 60 trillion tokens.
35:58
For the validation loss, you can also observe that it's
36:02
consistently lower than the original curve, demonstrating this
36:08
stability across optimisation, and also achieves the best improvement
36:14
on some of these coding, math, and reasoning heavy tasks,
36:18
as shown in the benchmark results of GPQA, Math, and HumanEval.
36:25
So the entire community keeps moving forward and we're happy that
36:30
we can, we're able to contribute to the community with new technologies
36:35
and some of these technologies have been sort of standard
36:41
and de facto for a long time, but as you can see, we still see a
36:45
lot of opportunities to improve it, to have revolutionary new design
36:52
to achieve better performance.
36:53
If you multiply all these scans together, you can actually
36:57
have a much better model.
36:59
So Adam was invented in 2014, and now we scale an
37:04
open-source MuonClip, a drop-in replacement for Adam.
37:09
And I'm sure that if you're training Transformer LLM,
37:14
it's going to be much better if you use MuonClip instead of Adam.
37:18
And Attention was invented over eight years ago.
37:22
And then now we have Kimi linear, which is a linear version.
37:25
We don't have to use full attention across all layers.
37:29
We can have linear attention that performs better on short,
37:32
long contacts at the same time.
37:35
And also, residual connections are now
37:39
Also a challenge, we scaled an open-source attention residue.
37:44
So I think one of the interesting things about our error is
37:49
that we sort of adopt a different mindset for doing research.
37:54
So if we go back to 10 years ago, it's mostly about publishing a
37:58
new idea, but then I think the lack of the rigor of the experiments
38:04
It's very hard to produce solid experimental results.
38:09
But now we have this scaling ladder.
38:10
We have enough resources to train the model and run it
38:16
at different scales.
38:17
We can have a whole set of benchmarks to measure the progress.
38:21
So it becomes easier to make confident and solid
38:26
conclusion out of it.
38:27
And this is one of the reasons why we are observing New
38:31
progress on this.
38:34
Ancient techniques, and I'm sure that we'll see more and
38:36
more, especially in the open source community.
38:40
I think we're going to have more and more even better
38:43
architectural and optimization improvement in the next few years.
38:50
All right, so to summarize, we're going to keep scaling our models.
38:54
And so these are three dimensions.
38:58
For example, we see different architectures and optimizers
39:04
that optimize all three dimensions.
39:07
And we'll keep. We see new dimensions for scaling, agent
39:12
swarms is not the end, and we are glad that we can move
39:16
forward with the entire open source community to achieve
39:20
better and better intelligence.
39:23
Thank you so much.
# How We Scaled Kimi K2.5
Zhilin Yang,Founder and CEO,Kimi (Moonshot AI)
AI-Generated Summary of this Video
Rate Now
Explore how we scaled Kimi K2.5 by pioneering the Muon optimizer to double token learning efficiency and maximize training throughput through Day 0 infrastructure co-design. Gain insights into AI-native training and linear attention architectures to help you unlock the potential of longer-running agents.
### Learn More About This Topic
Share
Favorite
Add to list
PDF
Events & Trainings:GTC San Jose
Date:March 2026
Industry:All Industries
Topic:Developer Tools & Techniques - Large Language Models (LLMs)
Level:General Interest
Language:English
NVIDIA technology:Blackwell, Hopper
Region:
Company Information
* [About Us](https://www.nvidia.com/en-us/about-nvidia/)
* [Investors](https://investor.nvidia.com/home/default.aspx)
* [Venture Capital (NVentures)](https://www.nvidia.com/en-us/startups/nventures/)
* [NVIDIA Foundation](https://www.nvidia.com/en-us/foundation/)
* [Research](https://www.nvidia.com/en-us/research/)
* [Corporate Sustainability](https://www.nvidia.com/en-us/sustainability/)
* [Technologies](https://www.nvidia.com/en-us/technologies/)
* [Careers](https://www.nvidia.com/en-us/about-nvidia/careers/)
News and Events
* [Newsroom](https://nvidianews.nvidia.com/)
* [Company Blog](https://blogs.nvidia.com/)
* [Technical Blog](https://developer.nvidia.com/blog/)
* [Webinars](https://www.nvidia.com/en-us/about-nvidia/webinar-portal/)
* [Stay Informed](https://www.nvidia.com/en-us/preferences/email-signup/)
* [Events Calendar](https://www.nvidia.com/en-us/events/)
* [GTC AI Conference](https://www.nvidia.com/gtc/events/)
* [NVIDIA On-Demand](https://www.nvidia.com/en-us/on-demand/)
Popular Links
* [Developers](https://developer.nvidia.com/)
* [Partners](https://www.nvidia.com/en-us/about-nvidia/partners/)
* [Executive Insights](https://www.nvidia.com/en-us/executive-insights/)
* [Startups and VCs](https://www.nvidia.com/en-us/startups/)
* [Documentation](https://docs.nvidia.com/)
* [Technical Training](https://www.nvidia.com/en-us/learn/organizations/)
* [Professional Services for Data Science](https://www.nvidia.com/en-us/support/enterprise/advisory-services/)
Follow NVIDIA
[](https://www.facebook.com/NVIDIA "")[](https://www.instagram.com/nvidia/?hl=en)[](https://www.linkedin.com/company/nvidia/)[](https://twitter.com/nvidia "")[](https://www.youtube.com/user/nvidia)
[United States](https://www.nvidia.com/en-us/location-selector/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [Your Privacy Choices](https://www.nvidia.com/en-us/about-nvidia/privacy-center/)
* [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/)
* [Accessibility](https://www.nvidia.com/en-us/about-nvidia/accessibility/)
* [Corporate Policies](https://www.nvidia.com/en-us/about-nvidia/company-policies/)
* [Product Security](https://www.nvidia.com/en-us/product-security/)
* [Contact](https://www.nvidia.com/en-us/contact/)
Copyright © 2026 NVIDIA Corporation
Select Location
The Americas
* [Argentina](https://www.nvidia.com/es-la/ "Argentina")
* [Brasil (Brazil)](https://www.nvidia.com/pt-br/ "Brasil (Brazil)")
* [Canada](https://www.nvidia.com/en-us/ "Canada")
* [Chile](https://www.nvidia.com/es-la/ "Chile")
* [Colombia](https://www.nvidia.com/es-la/ "Colombia")
* [México (Mexico)](https://www.nvidia.com/es-la/ "México (Mexico)")
* [Peru](https://www.nvidia.com/es-la/ "Peru")
* [United States](https://www.nvidia.com/en-us/ "United States")
Europe
* [België (Belgium)](https://www.nvidia.com/nl-nl/ "België (Belgium)")
* [Belgique (Belgium)](https://www.nvidia.com/fr-be/ "Belgique (Belgium)")
* [Česká Republika (Czech Republic)](https://www.nvidia.com/cs-cz/ "Česká Republika (Czech Republic)")
* [Danmark (Denmark)](https://www.nvidia.com/da-dk/ "Danmark (Denmark)")
* [Deutschland (Germany)](https://www.nvidia.com/de-de/ "Deutschland (Germany)")
* [España (Spain)](https://www.nvidia.com/es-es/ "España (Spain)")
* [France](https://www.nvidia.com/fr-fr/ "France")
* [Italia (Italy)](https://www.nvidia.com/it-it/ "Italia (Italy)")
* [Nederland (Netherlands)](https://www.nvidia.com/nl-nl/ "Nederland (Netherlands)")
* [Norge (Norway)](https://www.nvidia.com/nb-no/ "Norge (Norway)")
* [Österreich (Austria)](https://www.nvidia.com/de-at/ "Österreich (Austria)")
* [Polska (Poland)](https://www.nvidia.com/pl-pl/ "Polska (Poland)")
* [România (Romania)](https://www.nvidia.com/ro-ro/ "România (Romania)")
* [Suomi (Finland)](https://www.nvidia.com/fi-fi/ "Suomi (Finland)")
* [Sverige (Sweden)](https://www.nvidia.com/sv-se/ "Sverige (Sweden)")
* [Türkiye (Turkey)](https://www.nvidia.com/tr-tr/ "Türkiye (Turkey)")
* [United Kingdom](https://www.nvidia.com/en-gb/ "United Kingdom")
* [Rest of Europe](https://www.nvidia.com/en-eu/ "Rest of Europe")
Asia
* [Australia](https://www.nvidia.com/en-au/ "Australia")
* [中国大陆 (Mainland China)](https://www.nvidia.com/zh-cn/ "中国大陆 (Mainland China)")
* [India](https://www.nvidia.com/en-in/ "India")
* [日本 (Japan)](https://www.nvidia.com/ja-jp/ "日本 (Japan)")
* [대한민국 (South Korea)](https://www.nvidia.com/ko-kr/ "대한민국 (South Korea)")
* [Singapore](https://www.nvidia.com/en-sg/ "Singapore")
* [台灣 (Taiwan)](https://www.nvidia.com/zh-tw/ "台灣 (Taiwan)")
Middle East
* [Middle East](https://www.nvidia.com/en-me/ "Middle East")
NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking "Accept All", you consent to our use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/). You can manage your cookie settings by clicking on "Manage Settings." By continuing to use this site or by clicking one of the buttons below, you agree to our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. To opt out of non-cookie personal information "sales" / "sharing" for targeted advertising purposes, please visit the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies. You can manage these settings in the NVIDIA [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
Manage Settings
Reject Optional Accept All
Cookie Settings
We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.
Under certain privacy laws, you have the right to direct us not to "sell" or "share" your personal information for targeted advertising. To opt-out of the "sale" and "sharing" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the "sale" and "sharing" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/).
Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking "Save and Accept" or "Decline All" at the bottom, you consent to the use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) in accordance with your settings and accept our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). For more information about our privacy practices, please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/).
Required Cookies
Always Active
These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.
Cookies Details
Performance Cookies
- [x] Performance Cookies
These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.
Cookies Details
Personalization Cookies
- [x] Personalization Cookies
These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.
Cookies Details
Advertising Cookies
- [x] Advertising Cookies
These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.
Cookies Details
Cookie List
Clear
* - [x] checkbox label label
Apply Cancel
Consent Leg.Interest
- [x] checkbox label label
- [x] checkbox label label
- [x] checkbox label label
Decline All Save and Accept
[](https://www.onetrust.com/solutions/consent-and-preferences/)
Copy debug info