Kimi K3 Fully Open-Sourced: Overseas Developers Claim It "Renders Anthropic's Trillion-Dollar Valuation Meaningless"

Chinese large language model company Moonshot AI has announced full open-sourcing, bringing the technical details behind K3 into the spotlight.

On July 27, Moonshot AI (Kimi) announced the release of Kimi K3's model weights and technical report, and open-sourced the key infrastructure technologies supporting Kimi K3 model training: MoonEP, Flash KDA, and Agent Env. The company aims to accelerate the deployment and popularization of frontier intelligence and promote AGI research.

According to Kimi, despite limited computational resources, K3 achieved a 2.5x improvement in model scaling efficiency through a series of technological innovations, including Kimi Delta Attention, Attention Residuals, and MoonEP. This means that, in terms of optimal computational efficiency, the intelligence output per unit of compute is approximately 2.5 times the original.

Previously, on July 16, Kimi launched Kimi K3 with 2.8 trillion parameters, becoming the largest open-source model globally by parameter count, instantly shaking the global tech community. Within 48 hours of its release, user requests approached the cluster's capacity limit. On the evening of July 19, Moonshot AI issued an announcement suspending new consumer subscriptions, prioritizing all computational power for existing paying users.

Currently, K3 has been fully open-sourced. According to the introduction, anyone can download and deploy the Kimi K3 model, whether for internal R&D or embedding into end-user products, free of charge. "We believe that for a technology with profound implications like AGI, a broad and open ecosystem is the most suitable soil to support it. To this end, we will continue to contribute our efforts," Kimi stated.

Overseas Media and Developers Praise K3

Regarding the open-sourcing of K3, foreign media analyzed that this poses a significant challenge to AI giants, particularly Anthropic and OpenAI. The latest models from these two companies are Claude Fable and the GPT series, respectively. K3's performance closely follows theirs, with operating costs appearing to be only 1/2 to 1/3 of theirs; if specific queries are cached, costs can even drop to one-tenth.

Other foreign media reported that US companies are rapidly adopting Kimi models, with Chinese AI models now accounting for 60% of total usage on platforms like OpenRouter. Compared to closed-source models, Kimi K3 also demonstrates superior cost-effectiveness, completing more tasks per dollar spent. Foreign media noted that the infrastructural nature of AI dictates that open-sourcing is necessary to generate widespread economic impact, as closed sources hinder application and customization, making open source ultimately more preferred.

Clement Delangue, CEO of Hugging Face, stated that K3 quickly topped the Hugging Face trending charts after its release, with a very strong community response.

Ahmad, Founder and CEO of North American AI infrastructure startup Osmantic AI, posted online, calling Kimi K3 the "Local Version of Fable 5" and stating, "Do you know why Anthropic hates open-source AI? Kimi K3 is free and downloadable, rendering their trillion-dollar valuation meaningless."

On July 28, according to a post by Design Arena on social media, Kimi K3 topped the "Python-PPTX Slides" arena with an Elo score of 1379, leading the second-place Claude Fable 5 (1303 points) by 76 points. The post referred to Kimi K3 as an "unrestricted model" and noted it secured first place with the "largest advantage" seen on the leaderboard.

Kimi founder Yang Zhilin has repeatedly expressed his views on open source publicly.

"Currently, closed-source models still occupy a portion of the market share, while open source continues to unlock more market space. Ultimately, if model capabilities reach equivalent levels, open source will be the absolute winner." In previous interviews with media outlets such as The Paper, Yang Zhilin stated that the core competitiveness of open source lies in building a win-win ecological cooperation model, where all participants benefit from technological iteration and industrial implementation, thereby spawning massive application scenarios and channels, forming a complete, self-circulating industrial ecosystem.

"The total market size of the open-source ecosystem will ultimately far exceed that of the closed-source model, which is currently the most important development strategy and opportunity for China's AI industry," Yang Zhilin mentioned.

What Are the Technical Highlights of K3

Reporters from The Paper reviewed the technical report released for Kimi K3 and identified three core highlights of this full open-sourcing:

First, this is the world's first open-source "3T-class" model.

The report shows that Kimi K3 has a total of 2.8 trillion (2.8T) parameters, with 104 billion parameters activated per response. It supports a 1 million token context window and natively supports image and video understanding. This is the first 3T-class model to release all weights openly; previously, open-source models mostly remained at the 1 trillion parameter level.

In terms of comprehensive performance, K3 only trails the two strongest closed-source models (Claude Fable 5, GPT-5.6 Sol) but consistently surpasses all other evaluated models, including Claude Opus 4.8 and GPT-5.5. This means that, following DeepSeek, Chinese companies have once again used "open source" to challenge the global frontier of AI standards.

Second, K3 demonstrates AI's "self-evolution" capability.

The report mentions that during an uninterrupted 48-hour run without human intervention, K3 designed and verified a prototype AI inference chip (nano-KPU) using open-source EDA tools: integrating 1.46 million standard cells within a 4 square millimeter area, with simulated inference speed exceeding 8,700 tokens per second.

Additionally, K3 developed a compiler named MiniTriton from scratch, connecting the entire chain from programming language frontend to NVIDIA GPU low-level code generation. In its benchmarks, its geometric mean performance exceeded PyTorch's official two modes. The report also disclosed that an early version of K3 had already undertaken most of the company's internal GPU kernel optimization work during later development stages.

Third, ultra-high cost-performance ratio.

According to BrowseComp (web browsing agent test), K3 achieved the highest score of 91.2%, with a cost of only $2.03 per task. This is half the cost of GPT-5.6 Sol and about an order of magnitude cheaper than the Claude series.

In the programming test Kimi Code Bench 2.0, Kimi scored only 4 points lower than Claude Fable 5, with costs at just 38% of the latter. According to the knowledge work test GDPval-AAv2, the gap between Kimi and GPT-5.6 Sol was within 50 Elo points, with costs 13% lower.

Regarding structural innovation, the report disclosed that K3 adopts a KDA+AttnRes structure to replace traditional attention structures (such as Transformer). In this update, Kimi's self-developed Kimi Delta Attention (KDA) played a key role: KDA enables "shorthand-style" linear attention, keeping memory costs constant regardless of context length; the remaining 1 layer uses Gated MLA for "close reading of the full text," ensuring no loss of precision at critical points.

Redefining "Open Source" and "Business"

Beyond the technical aspects, the K3 license released by Kimi has also attracted widespread attention in the industry, featuring two key elements:

1. Companies operating Model as a Service (MaaS) businesses, with annual revenues exceeding $20 million for themselves and affiliated parties, need to sign separate commercial agreements.

2. If any commercial product or service has more than 100 million monthly active users, or monthly revenue exceeds $20 million, the "Kimi K3" label must be displayed prominently in the interface.

This implies that "open source" is not equivalent to "free." Kimi is attempting to define the balance between open accessibility and commercial returns.

Following Kimi's announcement of open-sourcing, major domestic cloud computing companies responded and adapted immediately.

Huawei Computing stated that upon the open-source release of Kimi K3, Ascend achieved Day 0 compatibility (referring to the ability to be fully compatible and operate normally on the "Day 0" of official product release or system launch): on the training side, training reproduction was completed on Atlas 800 A3 and Atlas 900 A3 SuperPoD based on MindSpeed MM, with specialized optimizations in operators, parallel acceleration, and video memory optimization; on the inference side, rapid deployment was achieved through vLLM Ascend and the SGLang open-source inference engine. Meanwhile, the Ascend 950 super node supports full series numerical precision formats such as FP8, MXFP8, and MXFP4, natively supporting Kimi K3's mxFP4 quantized weights.

Alibaba Cloud announced that the Zhenwu M890 super node instance has achieved Day 0 adaptation for Kimi K3, marking the first super node in China to successfully run a near-3-trillion-parameter model. This super node uses ICN Switch high-speed interconnect chips to fully interconnect 64 Pingtouge Zhenwu M890 chips at 800GB/s, with a video memory scale reaching 9TB. A single instance can support the expert parallelism requirements of large-scale trillion-level MoE models.

To improve operational efficiency, Alibaba Cloud, Pingtouge, and the Kimi team conducted deep joint adaptations at the software stack and operator levels. After optimization, the first token latency was reduced by approximately 35%, and single-card decoding throughput increased by 1.8 times, stably supporting 1M long contexts. The Qwen AI platform and Alibaba Cloud Bailian will simultaneously provide Kimi K3 model APIs. Alibaba Cloud stated that this adaptation verifies that domestic computational power possesses the capability to support large-scale applications of ultra-large parameter models.

Risk Warning and Disclaimer

The market involves risks; investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this content is at the user's own risk.