SemiAnalysis: Kimi K3 Surpasses NVIDIA's Open-Source Model, Committee Approach Falls Short
Claire Weston
Semiconductor research firm SemiAnalysis says Moonshot AI's Kimi K3 significantly outperformed Nvidia's flagship open-source model Nemotron 3 Ultra on benchmarks — and calls Jensen Huang's committee-driven development approach a proven dead end for American open-source AI.
Where exactly did Kimi K3 win?
Kimi K3 significantly outperformed Nvidia's flagship open-source model Nemotron 3 Ultra across benchmarks.
This means → a single Chinese independent lab has beaten the combined output of Nvidia's entire open-source coalition.
SemiAnalysis named Jensen Huang directly, arguing the Nemotron Committee model he championed is not the right path for U.S. open-source AI.
What went wrong with the committee approach?
SemiAnalysis calls the failure structural: the committee mechanism restricted free movement across different technical approaches, producing groupthink — consensus-seeking that suppresses dissent.
In plain terms = open source thrives on "anyone can experiment," while a committee enforces "everyone walks one road." The two are fundamentally incompatible.
The operational cracks showed too: Mistral made errors during Nemotron's pre-training phase, and under a single-committee architecture, one member's mistake dragged down the whole group, triggering cascading frustration among partners.
Why did the alliance turn on itself?
According to SemiAnalysis, some alliance members refused to share their best ideas, highest-quality datasets, and evaluation methods with the coalition.
This means → the alliance shared resources in name only; in practice, every party held back its best cards, hollowing out the foundation of cooperation from within.
This reflects a deeper contradiction: asking rival-tier companies to openly share core assets runs against basic incentive structures.
How does the Chinese AI lab ecosystem differ?
SemiAnalysis pointed to China's AI lab ecosystem as a contrast: labs compete with each other while freely borrowing proven innovations from one another.
This mechanism has produced distinct advances such as KDA, MSA, and DSA.
In plain terms = the Chinese model is not "sit at one table and agree," but "run your own track, and whatever works becomes fair game for everyone" — free-market competition drives iteration, not committee coordination.
What fix does SemiAnalysis propose for Nvidia?
SemiAnalysis is not calling on Nvidia to abandon the committee model entirely; it offers a specific structural remedy.
The core recommendation: build at least three fully isolated, non-communicating independent committees that innovate and compete separately across data, reinforcement learning, pre-training, and evaluation.
This means → replace "unified command" with "internal horse races," so the competitive mechanism operates inside the alliance too — whether Nvidia adjusts its open-source AI strategy accordingly will be the key test of whether it can close the gap.
Content is for reference only, not financial advice.