China’s Moonshot AI Unveils Kimi K3: The Largest Open-Source Model Poised to Compete with Leading U.S. Systems
Moonshot AI, an artificial intelligence startup located in Beijing and supported by Alibaba, has unveiled Kimi K3 — a groundbreaking model boasting 2.8 trillion parameters, which the company claims is the largest open-source AI model globally. Benchmarks indicate that it competes closely with top proprietary systems from major players like Anthropic and OpenAI.
This release comes just before the 2026 World Artificial Intelligence Conference in Shanghai, signaling a significant step forward in the global AI landscape and representing a pivotal moment for the open-source AI community. Additionally, it highlights a remarkable recovery for a business that faced setbacks in the past 18 months due to the rapid ascent of competitors like DeepSeek.
Complete model weights are expected to be available on July 27, as noted by researchers who have examined the company’s documentation. Those interested in experiencing Kimi K3 can do so immediately by signing up at kimi.com with a Google account or phone number, without needing a credit card, and engaging with what may be the most powerful open-source model developed to date.
Exploring the architecture behind the largest open-source AI model
Kimi K3 is classified as a frontier-class large language model, featuring 2.8 trillion parameters — around 75% more than DeepSeek’s V4 Pro, which has approximately 1.6 trillion parameters. It offers a 1-million-token context window, built-in visual comprehension, and an ongoing reasoning mode referred to as “thinking mode.”
This model incorporates two noteworthy innovations developed in-house at Moonshot AI: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which replaces traditional residual connections for enhanced scalability. Both techniques were previously shared as open research on GitHub.
In terms of API functionality, Kimi K3 is compatible with the OpenAI SDK, significantly easing integration for developers already utilizing OpenAI or Anthropic toolsets. Pricing is set at $3 per million input tokens and $15 per million output tokens, with cached input tokens available for just $0.30 per million, positioning it competitively against mid-tier options from Western labs while delivering performance close to the highest tier. A promotional rebate running until August 12 offers up to 30% back on API credits of $1,000 or more.
A representative from Moonshot AI elaborated on the parameter count, likening parameters to neural connections in the brain. With nearly 3 trillion of them, the model can “retain more knowledge and patterns, allowing for deeper understanding, more accurate thinking, and improved answers.”
Benchmark outcomes illustrate Kimi K3’s competition with Claude and GPT
On GDPval-AA v2, a benchmark assessing real-world tasks across 44 job categories and nine industries, Kimi K3 achieved a score of 1,687, ranking third overall, trailing only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), while outpacing Claude Opus 4.8 (1,600).
In the AA-Briefcase evaluation, which tests long-term knowledge work, K3 secured second place with a score of 1,527, surpassing GPT-5.6 Sol Max (1,495) and coming in behind only Fable 5 Max (1,587).
Remarkably, K3 earned a top score of 91.2 out of 100 on BrowseComp, a benchmark for high-difficulty information-seeking tasks.
The company claims this success is achievable with a single-agent setup utilizing its 1-million-token context window, without employing any context compression techniques, indicating that extended context management could prove more effective than complex multi-agent approaches.
One prominent AI analyst noted the shift in the open-source landscape on social media: “Open source is no longer trailing six months behind Western closed-source models. Reflect on that and consider its implications.”
This statement underscores the significance of Kimi K3’s launch, as open-source models have historically lagged behind their proprietary counterparts. Kimi K3 seems to have nearly eliminated that gap.
Ambitions evident in a 48-hour autonomous chip design demonstration
Beyond benchmark scores, Moonshot AI displayed a concept that highlights K3’s potential and the company’s strategic ambitions.
In a documented demonstration, Kimi K3 was given the task of designing a physical chip for a nano-scale version of itself. Over 48 hours, K3 effectively fulfilled the entire chip construction process, from design through optimization and verification, using open-source electronic design automation tools. This resulted in a functional tiny chip design, measuring just 4 square millimeters, achieving a timing convergence of 100 MHz and decoding over 8,700 tokens per second in simulation.
This chip is not intended for production but serves as evidence of Moonshot AI’s perspective on the future—advancing long-range autonomous agent capabilities. The requirement for coherent, multi-step work over an extended timeline reflects a major evolution from the straightforward question-answering models prevalent in the first generation of large language models.
The company also highlighted K3’s contributions to computational astrophysics, where it replicated a complex calculation, traditionally requiring an experienced researcher’s effort of one to two weeks, in around two hours. K3 did this by reviewing and cross-validating over 20 academic papers while implementing a comprehensive numerical pipeline.
The journey of Moonshot AI reflects the challenges of China’s competitive AI landscape
To grasp the significance of Kimi K3, it’s important to recall Moonshot AI’s standing 18 months ago and its rapid decline.
Founded in 2023 by Yang Zhilin, a Tsinghua University alumnus with prior research experience at Google and Meta, Moonshot AI quickly emerged among China’s most prominent AI startups. The company gained traction in 2024 as users flocked to its Kimi platform, notable for its long-text analysis and AI search capabilities. By early 2026, it had secured around $1.5 billion in funding across multiple rounds, with its valuation surging from $2.5 billion to $4.3 billion, and the company reportedly eyeing a new funding round at a valuation of $5 billion.
Then, in January 2025, DeepSeek launched a low-cost R1 model that disrupted the Chinese AI market, impacting Moonshot AI significantly. Once ranked third for monthly active users in China, Kimi plummeted to seventh place. In response, the company shifted to open-source models, beginning its pivot with Kimi K2 in July 2025 and further advancing with K2.5 in January 2026, largely aiming to re-establish its standing.
Kimi K3 is a culmination of this endeavor. The size of the model suggests that Moonshot AI has been strategically preparing for this moment for quite some time, as training a model of 2.8 trillion parameters necessitates significant computational resources and extensive planning well ahead of public release.
The geopolitical significance of releasing the largest open-source model
Releasing K3’s complete weights on July 27 carries important strategic implications.
A timeline chart from the company positions K3 as an extraordinary outlier, towering over competitors such as DeepSeek (1.6T), Xiaomi (1.02T), and Alibaba (397B). By launching the world’s largest open-source model, Moonshot AI hopes to establish itself as the central hub for the global open-source AI developer community.
This aligns with a broader trend among Chinese AI companies. Open-sourcing enables firms to showcase their technological prowess and broaden their developer communities and global influence—strategies aimed at countering U.S. restrictions on China’s tech progression. While several companies, including DeepSeek, Alibaba, Tencent, and Baidu, have introduced open-source models, none have matched this parameter count.
For enterprise technology leaders, the implications are substantial. A 2.8-trillion-parameter open-source model that approaches frontier performance opens new options for organizations looking to fine-tune, self-host, or build proprietary systems without being tethered to API agreements with OpenAI or Anthropic. However, running a model of this magnitude will require substantial GPU infrastructure, as inference at 2.8 trillion parameters cannot be handled by a single server.
Moonshot AI appears aware of this challenge, as evidenced by its Mooncake project, which received the Best Paper award at FAST 2025 for pioneering KV-cache-centric disaggregated serving for large language models — an architecture aimed specifically at enhancing inference scalability and cost-effectiveness.
A focus on Kimi Code and a three-tier model lineup for enterprise solutions
In tandem with K3, Moonshot AI is heavily investing in its coding agent ecosystem. Kimi Code, an open-source coding tool positioned against Anthropic’s Claude Code and Google’s Gemini CLI, was updated on K3’s launch day — with versions 0.25.0 and 0.26.0 adding features like enhanced subagent tools, background task management, and security enhancements.
The Kimi Code CLI has garnered over 3,100 stars on GitHub and offers integration with VSCode, Cursor, and Zed. The latest updates expanded the “coder subagent” capabilities to include background tasks, to-do lists, planning modes, skill invocation, and nested agents — effectively transforming the coding agent into a sophisticated system that can oversee complex software engineering projects with minimal human assistance.
This is strategically important, as coding tools are emerging as a critical revenue stream for AI labs. Anthropic disclosed that Claude Code reached an annualized recurring revenue of $1 billion in January. By constructing Kimi Code as an open-source alternative that centers on Kimi’s own models while also allowing other providers, Moonshot AI intends to capture developer workflows and, ultimately, enterprise contracts.
The current model lineup includes three tiers: K3 as the flagship ($3/$15 per million tokens for input/output), K2.7 Code specializing in coding tasks ($0.95/$4), and K2.6 as a general-use option ($0.95/$4). All three support context windows of 256,000 tokens or more, while K3 offers the full 1-million-token window. Context caching is automatic without the need for additional cache management, providing a developer convenience advantage over competitors requiring explicit cache control.
Kimi K3’s implications for the future of enterprise AI and the global model landscape
The introduction of Kimi K3 necessitates a reassessment of longstanding assumptions surrounding enterprise AI strategy.
The performance divide between open-source and proprietary models appears to have all but vanished at the frontier. If K3’s benchmark results are validated through independent evaluations and once the open weights become available for community testing on July 27, closed-source providers will find it challenging to justify their pricing based solely on capability.
Moreover, the center of AI innovation is shifting. In light of recent challenges with chip export regulations, China’s AI sector has now produced a model that effectively competes with top systems from firms enjoying direct access to advanced Nvidia hardware. The architectural innovations underpinning K3, especially the hybrid linear attention mechanism, indicate that algorithm efficiency is as critical as raw computing power.
Additionally, the diverse capabilities portrayed by K3 — including chip design and complex long-term research — herald a future where AI models won’t merely answer inquiries but autonomously undertake intricate projects over extended periods. This transition for enterprises regarding AI investments elevates the value proposition from “productivity assistant” to “autonomous technical workforce.”
Xinhua, the state news agency in China, defined the release as a national milestone, stating that K3 symbolizes progress in the development of China’s AI models. Liu Tieyan, dean at the Zhongguancun Academy in Beijing, remarked that a wave of Chinese open-source models is evolving from isolated successes to collective advancements, providing “new solutions and new pathways” for global AI development.
Just two years ago, Moonshot AI was an ambitious startup aiming to tackle daunting challenges. Eighteen months ago, it was a cautionary example of how quickly a market leader could falter. Today, it stands as the creator of the world’s largest open-source AI model — capable of designing a chip to run itself given a 48-hour window and internet access. The competitive landscape has shifted dramatically, and the race is now more intense than ever.
