China's Moonshot AI Unveils Kimi K3, a 2.8 Trillion-Parameter Open Model That Trades Blows With Claude and GPT

Moonshot AI dropped a big one on Thursday: Kimi K3, a 2.8-trillion-parameter language model the Beijing-based startup is calling the largest open-source AI system ever built. More striking than the size, though, is what the benchmarks say: K3 is landing punches against the best closed models Anthropic and OpenAI have to offer.

The timing isn't an accident. Moonshot AI, which is backed by Alibaba, released K3 just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, giving the model maximum visibility heading into one of the industry's biggest annual gatherings. It's also a remarkable reversal for a company that, as recently as 18 months ago, looked like it was fading into irrelevance.

You don't have to take Moonshot's word for any of this. The model is live now at kimi.com, sign up with a Google account or a phone number, no credit card needed. Full model weights are expected to follow on July 27, according to researchers who've reviewed the company's technical documentation. Once those weights are public, anyone with sufficient hardware can download and run what may be the most capable open model in existence.

What's actually inside K3

At 2.8 trillion total parameters, K3 is roughly 75 percent larger than DeepSeek's V4 Pro, which sits around 1.6 trillion parameters by Moonshot's own count. It ships with a 1-million-token context window, native visual understanding, and what the company calls "thinking mode," an always-on reasoning layer baked into the model rather than something toggled on for specific tasks.

The architecture leans on two innovations Moonshot developed internally and had already published as open research on GitHub: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which the company describes as a drop-in replacement for standard residual connections that produces more consistent gains as models scale up.

On the practical side, K3's API is compatible with the OpenAI SDK, so developers already building on OpenAI or Anthropic's tooling can switch over with minimal friction. Pricing lands at $3 per million input tokens and $15 per million output tokens, with cached inputs dropping to just $0.30 per million. That pricing puts K3 in line with mid-tier Western offerings while claiming near-top-tier performance. Moonshot is also running a promotion through August 12 that gives up to 30 percent back in vouchers on API credit purchases of $1,000 or more.

According to Xinhua, China's state news agency, a Moonshot AI executive described the parameter count in simple terms: parameters function something like neural connections in the human brain, and nearly 3 trillion of them means the model can store more knowledge, understand more, and answer with greater precision.

The benchmarks tell the real story

Specs are one thing. Performance is another, and this is where K3 gets genuinely interesting.

On GDPval-AA v2, a benchmark that tests real-world tasks spanning 44 occupations and nine industries, K3 scored 1,687, good for third place overall behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600). On AA-Briefcase, a private agentic benchmark from analytics firm Artificial Analysis built to test sustained, long-horizon knowledge work, K3 climbed to second place with a 1,527, edging out GPT-5.6 Sol Max (1,495) and falling just short of Fable 5 Max (1,587).

The standout result came on BrowseComp, a benchmark measuring difficult, long-horizon information retrieval, where K3 posted a state-of-the-art 91.2 out of 100. Moonshot says it hit that number in a single-agent setup, relying purely on the model's 1-million-token context window without any context compression or management tricks. That result suggests sheer context length, paired with strong retrieval, might outperform the more elaborate multi-agent setups other labs have leaned on.

One widely followed AI commentator captured the moment on social media: "Open source is no longer lagging six months behind Western closed-source models. Read that again, and think about what it all means."

That's the real headline buried in the benchmarks. For most of the last three years, open-source models have run a meaningful distance behind their closed counterparts. K3 looks like it's closed that distance almost entirely.

A chip that designs itself

The most striking demonstration Moonshot showed off wasn't a benchmark score at all. It was a 48-hour autonomous engineering exercise. The company tasked K3 with designing a physical chip capable of running a nano-scale version of itself. Left to work continuously as an autonomous agent for 48 hours, K3 carried the project through its full pipeline, from architecture through optimization and verification, using open-source electronic design automation tools, with no human intervention along the way.

The result wasn't a production-ready chip, but it was a real, functional design: just 4 square millimeters, achieving timing convergence at 100 MHz and capable of decoding more than 8,700 tokens per second in simulation.

That matters less as a chip-design milestone than as a signal of where Moonshot thinks the competition is headed next: long-range autonomous agents that can sustain coherent, multi-step technical work, reading documentation, making design calls, running verification loops, and recovering from failures over days rather than minutes. In a separate example, the company said K3 reproduced the universal I-Love-Q relation, a complex calculation in computational astrophysics that typically takes a senior researcher one to two weeks, in roughly two hours, reading and cross-checking more than 20 academic papers and building a full numerical pipeline in the process.

From also-ran to frontier lab

To appreciate why this release matters, it helps to know how far Moonshot AI had fallen. Founded in 2023 by Yang Zhilin, a Tsinghua University graduate who'd previously done research stints at Google and Meta, Moonshot quickly became one of the most closely watched AI startups in China. Its Kimi platform caught on in 2024 for long-text analysis and AI-powered search, and the company raised roughly $1.5 billion across multiple funding rounds, with its valuation jumping from $2.5 billion to $4.3 billion. By early 2026, it was reportedly seeking a new round that would value it at $5 billion.

Then DeepSeek's R1 model landed in January 2025 and scrambled the entire Chinese AI market. Moonshot was among the hardest hit. Kimi fell from third to seventh place in monthly active users in China almost overnight. The company's response was to go open-source, starting with Kimi K2 in July 2025 and accelerating with K2.5 in January 2026. K3 is the payoff from that bet. Its sheer scale suggests Moonshot had been building toward this release for a long time: training a model with 2.8 trillion parameters takes months of infrastructure work and planning, decisions that would have been locked in well before anyone outside the company saw the results.

Why open-sourcing the biggest model is also a geopolitical move

The decision to release K3's full weights on July 27 carries weight beyond pure technical bragging rights. Moonshot's own published timeline of open-source frontier models shows K3 towering over rivals: DeepSeek at 1.6 trillion parameters, Xiaomi at 1.02 trillion, Alibaba at 397 billion. By putting out the largest open model in the world, Moonshot is making a direct play to become the gravitational center of the global open-source developer community.

It's part of a broader pattern. As Reuters has noted, Chinese AI companies use open-sourcing to showcase technical capability and build developer communities and global influence, a strategy that also helps China push back against U.S. efforts to restrict its tech progress. DeepSeek, Alibaba, Tencent, and Baidu have all released open models of their own, but none at anything close to this parameter count.

For enterprise buyers, the practical upside is real: a 2.8-trillion-parameter open model performing near the frontier gives companies the option to fine-tune, self-host, or build proprietary systems on top of a serious base model, without being locked into API contracts with OpenAI or Anthropic. The catch is that running inference on a model this size requires serious GPU infrastructure. This isn't something that fits on a single server rack. Moonshot has at least signaled it's thought about that problem: its Mooncake project, which won the Best Paper award at FAST 2025, pioneered a KV-cache-centric disaggregated serving architecture specifically designed to make inference at extreme scale more practical and affordable.

Building the developer ecosystem around it

Moonshot isn't just shipping a model. It's building the tooling around it. Kimi Code, the company's open-source coding agent that competes directly with Anthropic's Claude Code and Google's Gemini CLI, got two updates the same day K3 launched, versions 0.25.0 and 0.26.0, adding expanded subagent tooling, background task management, and security fixes. The Kimi Code CLI has racked up more than 3,100 stars on GitHub and plugs into VSCode, Cursor, and Zed. The newest release expands the coder subagent's toolkit to include background tasks, todo lists, plan mode, skill invocation, and nested agents, turning it into a layered system capable of handling complex software projects with minimal hand-holding.

That investment isn't incidental. Coding tools have become one of the biggest revenue drivers in the AI industry. Anthropic disclosed in January that Claude Code alone had reached $1 billion in annualized recurring revenue. By offering an open-source coding agent that defaults to its own models but supports other providers, Moonshot is clearly angling to capture developer mindshare now and enterprise contracts later.

The company's lineup now spans three tiers: K3 as the flagship at $3/$15 per million tokens, K2.7 Code as a dedicated coding model at $0.95/$4, and K2.6 as a general-purpose option, also at $0.95/$4. All three support context windows of at least 256,000 tokens, with K3's full 1-million-token window at the top. Context caching happens automatically across the lineup, with no cache ID, TTL, or extra configuration required. It's a small but genuinely useful edge over competitors that make developers manage caching manually.

The bigger picture

If K3's benchmark numbers hold up once the open weights are out and the broader community gets to test them independently, it will be hard for closed-source providers to justify premium pricing on capability alone. The gap that's defined the open-versus-closed debate in AI for the past three years has, at least for this moment, mostly disappeared.

It also reinforces a point that's been building for a while: China's AI industry, despite real constraints on access to advanced chips, has produced a model that goes toe-to-toe with systems built by companies with direct access to Nvidia's best hardware. The hybrid attention mechanism underpinning K3 hints that clever architecture can substitute for raw compute in ways that weren't obvious a year or two ago.

Xinhua framed the release in explicitly national terms, calling it a milestone that "marks a new step forward in the development of China's artificial intelligence models." Liu Tieyan, dean of the Zhongguancun Academy in Beijing, told the outlet that China's wave of open-source models has shifted from scattered breakthroughs to something more collective, offering "new solutions and new paths" for AI development worldwide.

Two years ago, Moonshot AI was a startup named for an ambition it hadn't yet proven it could meet. Eighteen months ago, it looked like a company that had already missed its window. Now it's behind the largest open-source AI model ever built, one that, given 48 hours and an internet connection, can design a chip capable of running itself. Whether K3's numbers hold up under broader scrutiny once the weights go public will say a lot about how real this moment is. Either way, the race to the frontier just got a lot more crowded.