Kimi released the K3 model with 2.8 trillion parameters, setting a new record for open-source model scale. The model ranks first in the frontend arena, surpassing Fable 5, and demonstrated capabilities such as autonomous chip design and GPU compiler development within 48 hours.
SynthePulse Insight · AI deep reading
Version 2 · 5 sources
On July 17, Kimi released the K3 large model, refreshing the open-source parameter ceiling with a scale of 2.8 trillion, and topped the Arena frontend code competition leaderboard with a score of 1679, surpassing Anthropic's Claude Fable 5. Meanwhile, its API prices were raised by nearly four times to benchmark against Sonnet 5, while co-founders explicitly stated they are not currently pursuing the video generation track.
On July 17, Kimi officially launched its new generation large model, K3. According to reports from Quantum Bit, the model possesses 2.8 trillion parameters, a scale that refreshes the current parameter upper limit for open-source models. Titanium Media AGI further pointed out that in evaluations released by Arena, K3 achieved an excellent score of 1679 on the frontend code competition leaderboard, successfully taking the top spot and surpassing Anthropic's Claude Fable 5. This result indicates that K3's coding capabilities in specific vertical domains have reached industry-leading levels.
Accompanying claims of technical capability, Kimi adopted an aggressive strategy in commercial pricing. Both InfoQ and AI Frontline reported that K3's API prices were increased by nearly four times, with the target clearly pointing to OpenAI's Sonnet 5. This significant price hike suggests that Kimi attempts to match its claimed high-end performance positioning through higher unit prices, or to cover substantial computing costs. However, specific details of the high pricing were not disclosed in the sources, only confirming the multiple relationship.
Against the backdrop of multimodal becoming the mainstream trend, Kimi chose a differentiated path. Co-founders explicitly stated that "video generation is difficult to improve intelligence," therefore K3 will not pursue the video generation field for now. This strategic judgment concentrates resources on core capabilities such as text, code, and logical reasoning. Although no specific experimental data was provided to support this view, it constitutes the core constraint condition of K3's product definition.
To enhance the appeal of the developer ecosystem, Kimi announced that K3's weights would be opened within ten days. Additionally, Quantum Bit mentioned that the team had previously demonstrated the ability to "autonomously build chips in 48 hours and write GPU compilers from scratch." While this is not a direct function of the K3 model itself, it reflects their technical reserves in underlying hardware collaborative optimization. These actions collectively construct Kimi's full-stack capability narrative, spanning from models to infrastructure.
This article's information is primarily based on reports from multiple technology media (AI Frontline, Quantum Bit, Titanium Media AGI, InfoQ, and Jiqizhixin). Among them, benchmark scores and release dates are confirmed facts cross-verified by multiple parties; pricing multiples and strategic statements are source claims; the judgment on the effectiveness of video generation is a statement of opinion by the co-founders, unverified by independent third parties.
Kimi K3 establishes its technical status through ultra-large parameter scales and high scores in specific domains, while attempting to enter the high-end market through significant price hikes. Its strategy of abandoning video generation reflects a focus on core intelligence capabilities. Subsequent attention should be paid to community feedback after open-sourcing and actual implementation effects.
Primary report
Primary source
Also covered by 3 sources