Moonshot AI Unveils 2.8 Trillion-Parameter "Kimi K3," Promises Open Weights

On July 16, 2026, China’s AI lab Moonshot AI unveiled its large language model “Kimi K3.” The company describes the model as its “most advanced to date,” boasting 2.8 trillion parameters. The model is now available via the company’s website and API, with an open-weight release promised by July 27, 2026.

Moonshot AI refers to Kimi K3 as the “first open 3T-class model,” rounding up its parameter count to 3 trillion. This surpasses the 1.6 trillion parameters of DeepSeek’s v4 Pro, making it one of the largest open-weight models ever released.

Performance and Benchmarks

According to the company’s self-reported benchmarks, Kimi K3 outperforms Claude Opus 4.8 max and GPT-5.5 high in most categories, though it falls short of Claude Fable 5 and GPT-5.6 Sol.

Third-party evaluator Artificial Analysis reported that Kimi K3 achieved an overall Elo score of 1547 in their private long-term knowledge task evaluation, representing a 732-point improvement over its predecessor, Kimi K2.6. This score ranks just below Claude Fable 5.

The cost per task for Kimi K3 is $0.94, comparable to GPT-5.6 Sol’s $1.04, and nearly half the cost of Opus 4.8, which stands at $1.80. However, it remains more expensive than other open-weight models.

On the Artificial Analysis Intelligence Index, Kimi K3 demonstrated a 21% reduction in output token usage compared to K2.6, indicating improved inference efficiency.

Additionally, Kimi K3 secured the top spot on Arena.ai’s Frontend Code leaderboard, surpassing Claude Fable 5 in code generation performance.

Pricing and Strategy

The pricing for Kimi K3 is set at $3 per million input tokens and $15 per million output tokens, on par with Anthropic’s Claude Sonnet series. This represents the highest price range for a Chinese AI lab-developed model.

In comparison, the previous Kimi K2.6 model had input and output token costs of $0.95 and $4, respectively. The substantial price hike is accompanied by a more than twofold increase in parameter count, from 1 trillion to 2.8 trillion.

Simon Willison of Simon Willison’s Weblog accessed Kimi K3 via OpenRouter and tested it with an image generation prompt for an “SVG of a pelican riding a bicycle.” The test generated 16,658 tokens of output (13,241 of which were inference tokens) in response to 95 input tokens, costing $0.25. When the output SVG was re-input as an image, the cost was $0.006, resulting in a depiction of a pelican on a bicycle wearing a red scarf.

Significance of Open Weights

Moonshot AI’s decision to release Kimi K3 with open weights marks a significant development in the Chinese AI industry. Following in the footsteps of DeepSeek, which gained international acclaim for its open-weight strategy, Moonshot’s move is expected to have a notable impact.

With 2.8 trillion parameters, Kimi K3 is the largest open-weight model to date, well above DeepSeek’s 1.6 trillion-parameter v4 Pro. Kimi K3 demonstrates a level of performance that rivals even some closed models, suggesting it could provide a competitive edge in open settings.

However, the increase in parameter count also comes with higher inference costs. The output cost of Kimi K3 is approximately 3.75 times higher than that of K2.6, a direct consequence of its expanded scale. Developers and enterprise users will need to weigh these performance gains against the increased cost.

Impact on the Industry

The introduction of Kimi K3 is set to intensify competition among Chinese AI labs. Companies like DeepSeek, Moonshot, Alibaba’s Qwen series, and ByteDance’s Doubao are all vying for dominance in the large language model market, competing on both performance and pricing.

By positioning Kimi K3 as a high-priced model, Moonshot appears to be pursuing a premium branding strategy. Historically, Chinese AI models have expanded their market share by offering lower costs, but Kimi K3 emphasizes top-tier performance while aligning its pricing with international high-end models.

That said, the commitment to releasing open weights comes with its own challenges. While Moonshot has set a clear deadline of July 27, 2026, for the release, it remains uncertain whether this timeline will be met. Regulatory hurdles and export controls in China could potentially delay the release.

Editorial Opinion

In the short term, the open-weight release of Kimi K3, if realized by July 27, could make waves in the open-source AI community. While running a 2.8 trillion-parameter model locally is improbable, its immense value as a research tool and foundation for fine-tuning cannot be overstated. Its superior performance in code generation, surpassing Claude Fable 5, also opens up new possibilities in developer tools.

From a long-term perspective, Kimi K3 signals that Chinese AI labs are seriously entering the international high-performance model arena. By raising prices while emphasizing quality, Moonshot is pursuing a strategy that could prove sustainable. However, balancing the promises of open weights with the costs of high-end APIs presents a significant business challenge. The key will be how effectively Moonshot can differentiate itself from competitors like DeepSeek, which has opted for a low-cost, open-weight strategy.

From an editorial standpoint, the question must be asked: is the sheer scale of 2.8 trillion parameters truly necessary? The exponential increase in parameters results in greater computational demands, raising concerns about environmental impact and access costs. As smaller, more efficient models gain traction, the AI industry may need to reevaluate the value of an arms race focused solely on parameter count.

References

• “Kimi K3, and what we can still learn from the pelican benchmark”, by Simon Willison — Simon Willison’s Weblog, 2026-07-16T20:19:30.000Z (ARR)
• Source URL: https://simonwillison.net/2026/Jul/16/kimi-k3/#atom-entries