The top 10 most intelligent open-source models all use a mixture-of-experts architecture.
Kimi K2 Thinking, DeepSeek-R1, Mistral Large 3 and others run 10x faster to enable one-tenth the cost per token on NVIDIA GB200 NVL72.
A look under the hood of virtually any frontier model today will reveal a mixture-of-experts (MoE) model architecture that mimics the efficiency of the human brain.
Just as the brain activates specific regions based on the task, MoE models divide work among specialized experts, activating only the relevant ones for every AI token. This results in faster, more efficient token generation without a proportional increase in compute.
The industry has already recognized this advantage. On the independent Artificial Analysis (AA) leaderboard, the top 10 most intelligent open-source models use an MoE architecture, including DeepSeek AI's DeepSeek-R1, Moonshot AI's Kimi K2 Thinking, OpenAI's gpt-oss-120B and Mistral AI's Mistral Large 3.
However, scaling MoE models in production while delivering high performance and low cost per token is notoriously difficult. The extreme codesign of NVIDIA GB200 NVL72 systems combines hardware and software optimizations for maximum performance and efficiency, making it practical and straightforward to scale MoE models.
The Kimi K2 Thinking MoE model - ranked as the most intelligent open-source model on the AA leaderboard - sees a 10x performance leap on the NVIDIA GB200 NVL72 rack-scale system compared with NVIDIA HGX H200. This 10x performance leap, also seen on other MoE models such as DeepSeek-R1 and Mistral Large 3, enables one-tenth the cost per token and underscores why NVIDIA's full-stack inference platform is the key to unlocking their full potential.
What Is MoE, and Why Has It Become the Standard for Frontier Models? Until recently, the industry standard for building smarter AI was simply building bigger, dense models that use all of their model parameters - often hundreds of billions for today's most capable models - to generate every token. While powerful, this approach requires immense computing power and energy, making it challenging to scale.
Much like the human brain relies on specific regions to handle different cognitive tasks - whether processing language, recognizing objects or solving a math problem - MoE models comprise several specialized experts. For any given token, only the most relevant ones are activated by a router. This design means that even though the overall model may contain hundreds of billions of parameters, generating a token involves using only a small subset - often just tens of billions.
Like the human brain uses specific regions for different tasks, mixture-of-experts models use a router to select only the most relevant experts to generate every token. By selectively engaging only the experts that matter most, MoE models achieve higher intelligence and adaptability without a matching rise in computational cost. This makes them the foundation for efficient AI systems optimized for performance per dollar and per watt - generating significantly more intelligence for every unit of energy and capital invested.
Given these advantages, it is no surprise that MoE has rapidly become the architecture of choice for frontier models, adopted by over 60% of open-source AI model releases this year. Since early 2023, it's enabled a nearly 70x increase in model intelligence - pushing the limits of AI capability.
Since early 2025, nearly all leading frontier models use MoE designs. Our pioneering work with OSS mixture-of-experts architecture, starting with Mixtral 8x7B two years ago, ensures advanced intelligence is both accessible and sustainable for a broad range of applications, said Guillaume Lample, cofounder and chief scientist at Mistral AI. Mistral Large 3's MoE architecture enables us to scale AI systems to greater performance and efficiency while dramatically lowering energy and compute demands.
Overcoming MoE Scaling Bottlenecks With Extreme Codesign Frontier MoE models are simply too large and complex to be deployed on a single GPU. To run them, experts must be distributed across multiple GPUs, a technique called expert parallelism. Even on powerful platforms such as the NVIDIA H200, deploying MoE models involves bottlenecks such as:
Memory limitations: For each token, GPUs must dynamically load the selected experts' parameters from high-bandwidth memory, causing frequent heavy pressure on memory bandwidth.
Latency: Experts must execute a near-instantaneous all-to-all communication pattern to exchange information and form a final, complete answer. However, on H200, spreading experts across more than eight GPUs requires them to communicate over higher-latency scale-out networking, limiting the benefits of expert parallelism.
The solution: extreme codesign.
NVIDIA GB200 NVL72 is a rack-scale system with 72 NVIDIA Blackwell GPUs working together as if they were one, delivering 1.4 exaflops of AI performance and 30TB of fast shared memory. The 72 GPUs are connected using NVLink Switch into a single, massive NVLink interconnect fabric, which allows every GPU to communicate with each other with 130 TB/s of NVLink connectivity.
MoE models can tap into this design to scale expert parallelism far beyond previous limits - distributing the experts across a much larger set of up to 72 GPUs.
This architectural approach directly resolves MoE scaling bottlenecks by:
Reducing the number of experts per GPU: Distributing experts across up to 72 GPUs reduces the number of experts per GPU, minimizing parameter-loading pressure on each GPU's high-bandwidth memory. Fewer experts per GPU also frees up memory space, allowing each GPU to serve more concurrent users and support longer input lengths.
Accelerating expert communication: Experts spread across GPUs can com
Most recent headlines
05/01/2027
Worlds first 802.15.4ab-UWB chip verified by Calterah and Rohde & Schwarz to be ...
04/08/2026
Dalet, a leading technology and service provider for media-rich organizations, t...
04/07/2026
April 7 2026, 19:00 (PDT) Detective Conan: Fallen Angel of the Highway Opens in...
01/06/2026
January 6 2026, 05:30 (PST) Dolby Sets the New Standard for Premium Entertainment at CES 2026
Throughout the week, Dolby brings to life the latest innovatio...
13/05/2026
New Adobe Premiere Color Grading Mode Accelerated on NVIDIA GPUs
Joel Pennington May 13, 2026
0 Comments
New NVIDIA RTX-accelerated features streamlin...
13/05/2026
Grass Valley announced that dB Broadcast has delivered new IP-based outside broadcast (OB) trucks for Cloudbass, featuring Grass Valley LDX 100 Series cameras a...
13/05/2026
Ikegami will exhibit the latest additions to its wide range of broadcast production cameras, control units, viewfinders and monitors on stand 5D3-1 at Broadcast...
13/05/2026
FISE, working with the founding members of the XR Sports Alliance (XRSA), Accedo, Qualcomm Technologies, Inc. and HBS, have collaborated to develop an immersive...
13/05/2026
Canon Unveils New EOS R6 V Full-Frame EOS Camera and RF20-50mm F4 L IS USM PZ Bu...
13/05/2026
Boston Conservatory at Berklee Honors Beth Morrison and Moses Pendleton at Comme...
13/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/05/2026
Creative software developer Foundry today announced the latest developments on Nuke Stage. A purpose-built application for end-to-end virtual production and in-...
13/05/2026
A definitive portrait of one of Ireland's most influential musicians
New TV documentary airs Monday 18 May on RT One and RT Player at 9.35pm
Watch the...
13/05/2026
Agentic AI is changing the way users get work done. Following the success of OpenClaw, the community is embracing new open source agentic frameworks. The latest...
13/05/2026
Reinforcement-learning agents - AI systems that learn by trial and error - can c...
12/05/2026
Beyond the Hype: A Strategic Post-Hoc Analysis of NAB 2026 If NAB Show 2026 had an underlying theme, it was a quiet, industry-wide pivot from the high-energy sp...
12/05/2026
Guntermann and Drunck (G&D), a Panoptec Technologies Group company, and CT Square, led by Chandresh Shah, have announced a joint venture to distribute G&D and V...
12/05/2026
With 30 days until the start of the FIFA World Cup 2026, Telemundo, the exclusive Spanish-language home of the tournament in the United States, has announced th...
12/05/2026
The NHL has announced the return of Stanley Pup for its third consecutive year, a 90-minute special featuring adoptable rescue dogs competing on a miniature rin...
12/05/2026
NBCUniversal presented its 2026 Upfront to advertisers at Radio City Music Hall, detailing upcoming programming across NBC, Peacock, Bravo, and Versant properti...
12/05/2026
FOX Sports has announced funding for the Fandom and Social Connection Initiative at Harvard Kennedy School's Shorenstein Center on Media, Politics, and Publ...
12/05/2026
TNDV and Live Media, both divisions of Live Media Group, supported live broadcast coverage around NCAA Final Four weekend in Indianapolis, including the March M...
12/05/2026
The European Football Alliance (EFA) has announced a content distribution agreement with Fubo Sports Network, the free ad-supported streaming TV (FAST) channel ...
12/05/2026
For the first time, Spanish-speaking fans in the U.S. will have two separate tel...
12/05/2026
CP Communications led a comprehensive spectrum management initiative on behalf of Churchill Downs during Kentucky Derby week, coordinating RF assets across the ...
12/05/2026
LiveU has announced a strategic partnership with DRONERESPONDERS, a 501(c)3 non-...
12/05/2026
Open Broadcast Systems has announced that BMC TV, a specialist in IP transport of broadcast content, has selected the Open Broadcast Systems 5G Flyaway solution...
12/05/2026
NEP Europe, part of NEP Group, has announced it will deliver broadcast solutions...
12/05/2026
Grass Valley has announced continued collaboration with Ravensbourne University ...
12/05/2026
Stats Perform has announced the launch of Opta Pulse, an AI-assisted video creation and distribution platform for leagues, rights holders, and broadcasters. The...
12/05/2026
FOX Sports has announced a collaboration with Sesame Workshop to integrate Sesame Street characters into FOX Sports' FIFA World Cup 2026 programming. Conten...
12/05/2026
To date, NHL Productions has produced 19 broadcasts with commentary in American Sign Language
NHL in ASL (American Sign Language) may be just one show, but the...
12/05/2026
Google's Brian Albert: creators, athletes, highlights, nostalgia, second-scr...
12/05/2026
A still from Past Lives by Celine Song, an official selection of the Premieres program at the 2023 Sundance Film Festival. (Courtesy of Sundance Institute | p...
12/05/2026
Spotify is where fans and artists come together, turning discovery into somethin...
12/05/2026
Features patented Marco-MMC clocking technology
Black Lion Audio's latest release combines the company's expertise in clocking with their renowned p...
12/05/2026
New Track Panel, sequencer upgrades & more
Following their recent public beta release, Reason Studios have announced the full release of Reason 14. With the...
12/05/2026
One month to go! SBS reveals expansive FIFA World Cup 2026 lineup beyond the pi...
12/05/2026
Rohde & Schwarz presents its advanced solutions for power electronics testing at...
12/05/2026
aconnic AG (ISIN: DE000A0LBKW6), Munich, has developed a modified fund raising p...
12/05/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/05/2026
Tyrell Corporation, specialists in high-end live sports and entertainment broadcasts, was tasked with delivering compelling broadcast coverage of premier equest...
12/05/2026
Registration is now open for IBC2026 as the global media, entertainment and technology community prepares to converge on the RAI Amsterdam from 11 14 September ...
12/05/2026
Ross Video, a global leader in live video production technology, will present its latest innovations and integrated production workflows at BroadcastAsia 2026, ...
12/05/2026
500 selected leaders from around the world across start-ups, corporates, and venture capital. Over 50bn in assets under management among attending investors, a...
The top 10 most intelligent open-source models all use a mixture-of-experts architecture.
