Moonshot AI launched Kimi K3 on July 16, and the specifications demand attention. According to the company's technical blog, K3 packs 2.8 trillion parameters into a Mixture-of-Experts architecture, making it the largest open-source model released to date. That's not a typo. It's nearly double the parameter count of DeepSeek's V4-Pro and roughly 70% larger than anything else from a Chinese lab this year.

The model uses what Moonshot calls Kimi Delta Attention and Attention Residuals, architectural innovations the company says improve information flow in long sequences and deep networks. K3 natively supports vision understanding and comes with a 1-million-token context window. For perspective: the previous K2.6 topped out at 256K tokens. This is a fourfold expansion in how much information the model can hold in working memory at once.

Benchmarks That Matter

Moonshot is citing numbers from Artificial Analysis. K3 scored 1687 on GDPval-AA v2, a leaderboard testing real tasks across 44 occupations in nine industries. That places it behind Claude Fable 5 Max and GPT-5.6 Sol Max but ahead of Claude Opus 4.8 Max, which scored 1600. On AA-Briefcase, a private benchmark measuring long-horizon agentic knowledge work, K3 scored 1527. Second place overall, ahead of GPT-5.6 Sol Max at 1495.

Advertisement

The 1-million-token context shows up in the BrowseComp result too. Moonshot says K3 hit a state-of-the-art score of 91.2 on the benchmark running as a single agent, with no context compression or workarounds needed to manage the window.

The Pricing Equation

Here's where things get interesting for anyone paying attention to AI economics. As TechCrunch reported, Kimi K3 costs $0.30 per million tokens for cached input, $3 per million tokens for uncached input, and $15 per million tokens for output. Compare that to Claude Opus 4.8 at $5 per million input tokens and $25 per million output. The Anthropic model is roughly 1.7x more expensive on input and nearly double on output.

For context, Moonshot's previous K2.6 model launched at $0.60 input and $2.50 output per million tokens. Third-party analysis at the time described that as roughly 8x cheaper on input and 10x cheaper on output than Claude Opus 4.7. K3 maintains aggressive pricing while delivering what the company claims is frontier-level performance.

Why This Changes the Calculus

The AI industry has operated under a comfortable assumption: the best models come from OpenAI and Anthropic, and you pay premium prices for premium capabilities. Chinese labs have been playing catch-up, typically trailing American counterparts by 8-12 months on capability benchmarks.

K3 compresses that gap to something that looks more like weeks than months. Early testers on X are comparing it directly to Fable 5 and GPT-5.6. One tester noted that on a frontend coding prompt, K3 produced output better than many frontier models, though it took 35 minutes to finish. Speed remains a trade-off.

Advertisement

Moonshot says the full model weights will be released in the coming days. If that happens, developers gain an open-weight option at frontier scale, something that didn't exist a week ago. The modified MIT license the company used for K2 allowed commercial use with minimal restrictions. The AI Race Just Got More Complicated when Thinking Machines released Inkling. K3 is a larger bet on the same thesis: open weights plus aggressive pricing can undercut the closed-model business model.

What Remains Unverified

Independent benchmarks are still thin. Most comparisons rely on community testing and Moonshot's self-reported numbers. The company has not published Terminal-Bench 2.1 scores, which would offer a cleaner apples-to-apples comparison with Anthropic's latest models. Early tester reports suggest K3 excels at long-horizon agent tasks and 3D generation but is notably slow on complex workloads.

Moonshot is also raising fresh capital at a $31.5 billion valuation, up from $20 billion in May. The company's annualized recurring revenue reportedly surpassed $300 million by mid-June. The funding suggests confidence in the K3 bet, though it also reflects the capital-intensive reality of training trillion-parameter models.

For enterprise teams evaluating AI infrastructure, the question is no longer whether open models can compete with closed ones. The question is whether the reliability, safety guarantees, and government-vetted access of Western labs justify paying 2-3x more per token. For many workloads, the answer may be shifting.