Moonshot AI, the Beijing-based startup backed by Alibaba, has released Kimi K3 — a model the company claims is the largest open-source AI ever built, with 2.8 trillion parameters and benchmark scores that rival the top proprietary systems from Anthropic and OpenAI.

Full model weights are scheduled to drop on July 27. In the meantime, the model is already accessible at kimi.com — no credit card required, just a Google account or phone number.

Architecture: What Makes K3 Different

At 2.8 trillion parameters, Kimi K3 is roughly 75% larger than DeepSeek V4 Pro's estimated 1.6 trillion parameters — the previous scale leader among open models.

The architecture introduces two internally developed innovations:

  • Kimi Delta Attention — a hybrid linear attention mechanism
  • Attention Residuals — described as a drop-in replacement for residual connections that delivers consistent scaling gains

Both techniques were previously published as open research on GitHub. The model also features a 1-million-token context window, native visual understanding, and an always-on reasoning mode the company calls "thinking mode."

For developers, K3 is compatible with the OpenAI SDK, meaning teams already building on OpenAI or Anthropic toolchains can integrate it with minimal friction. Pricing sits at $3 per million input tokens and $15 per million output tokens, with cached input dropping to $0.30 per million — competitive with mid-tier Western offerings, but positioned at claimed top-of-market performance.

Benchmark Results: Trading Blows With Claude and GPT-5

The numbers, drawn from public leaderboards and a private evaluation by analytics firm Artificial Analysis, are striking.

On GDPval-AA v2 — a real-world task benchmark spanning 44 occupations and 9 industries — K3 scored 1,687, placing third 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, Artificial Analysis's private agentic benchmark for long-horizon knowledge work, K3 came in second with 1,527, beating GPT-5.6 Sol Max (1,495) and trailing only Fable 5 Max (1,587).

Most notably, K3 achieved 91.2 out of 100 on BrowseComp, a benchmark for difficult long-horizon information retrieval — accomplished in a single-agent setup using its full 1-million-token context window, with no compression or multi-agent scaffolding.

"Open source is no longer lagging six months behind Western closed-source models. Read that again, and think about what it all means."

That observation, widely circulated in AI research circles, captures the inflection point this release represents.

The 48-Hour Autonomous Chip Design Demo

Beyond benchmarks, Moonshot AI demonstrated something arguably more revealing of K3's strategic intent.

Over a 48-hour autonomous agent run, K3 was tasked with designing a physical chip capable of running a nano-scale version of itself. Working independently using open-source electronic design automation tools, the model completed the full pipeline — from architectural design through optimization and verification — producing a 4-square-millimeter chip design that achieved timing convergence at 100 MHz and could decode more than 8,700 tokens per second in simulation.

This is not a production chip. It's a proof of concept for sustained, multi-step autonomous technical work — the kind of long-range agentic capability that represents the next competitive frontier for foundation model labs.

The company also highlighted a computational astrophysics case where K3 reproduced the I-Love-Q relation — a calculation that typically takes a senior researcher one to two weeks — in approximately two hours, reading and cross-validating more than 20 papers and implementing a complete numerical pipeline.

Moonshot's Fall and Comeback

To appreciate the significance of K3, you need to understand where Moonshot AI stood 18 months ago.

Founded in 2023 by Yang Zhilin — a Tsinghua University graduate with prior research stints at Google and Meta — Moonshot gained early traction in 2024 through its Kimi platform's long-text analysis and AI search capabilities. By early 2026, the company had raised approximately $1.5 billion, with its valuation climbing from $2.5 billion to $4.3 billion and a reported new round targeting $5 billion.

Then DeepSeek happened. DeepSeek's low-cost R1 model, released in January 2025, reshaped China's AI market almost overnight. Kimi, which had ranked third in monthly active users in China, fell to seventh. Moonshot's pivot to open-source — beginning with Kimi K2 in July 2025 and K2.5 in January 2026 — was a calculated effort to rebuild relevance.

Kimi K3 is the culmination of that strategy. Training a 2.8-trillion-parameter model requires enormous compute and months of preparation, which means the architectural decisions behind K3 were locked in well before today's release.

What This Means for Developers and Founders

The practical implications of K3 are worth spelling out directly:

  • Open-source parity with proprietary frontier models is no longer a future aspiration — it appears to have arrived.
  • Teams that have avoided open-source for performance reasons now have a credible alternative at frontier scale, with the cost and customization advantages open weights bring.
  • The OpenAI SDK compatibility means migration friction is low for teams already in that ecosystem.
  • The geopolitical dimension — a Chinese lab open-sourcing the world's largest model — will accelerate pressure on U.S. labs to respond, either by releasing their own weights or by differentiating more aggressively on safety, tooling, and enterprise integration.

For startup founders building AI-native products, the message is clear: the open-source model landscape has fundamentally shifted, and the cost-performance calculus for proprietary APIs just got a lot more competitive.