Moonshot AI’s Kimi K3 tops Claude Fable 5 in writing benchmark

Moonshot AI has released Kimi K3, a new Chinese open-weight AI model that ranked ahead of Anthropic’s Claude Fable 5 on a specialized writing benchmark. For investors tracking the AI race, the key point is access: open-weight models can be run, modified or built on outside the company that created them once the model files are available.

The company introduced Kimi K3 on July 16 as a 2.8 trillion-parameter model. Parameters are the internal settings a model learns during training, and a higher count often signals more capacity, though it does not guarantee better performance in every task.

Moonshot AI said Kimi K3 supports a 1 million-token context window, meaning it can process very large amounts of text or data in a single prompt. A token is a chunk of text used by AI systems for reading, generating and pricing output. The company also described Kimi K3 as native multimodal, meaning it is built to handle more than one type of input.

Where Kimi K3 scored well

Towards AI’s Writing Elo benchmark placed Kimi K3 at 2,840, above Claude Fable 5’s listed maximum score of 2,760. The benchmark asks models to write scripts that are judged blind against published versions, then scores them using Elo, a ranking system best known from chess.

Louis-François Bouchard of Towards AI said on X that early internal results put Kimi K3 at No. 1 for writing in the group’s editorial voice. He said that marked a move from No. 21 for Kimi K2.6 to No. 1 for Kimi K3, and added that the model runs at about $0.25 per script.

Benchmarks measure specific tasks, so the result does not say Kimi K3 is better than every rival model across all uses. It does show that Moonshot’s newest model performed strongly on a test focused on script writing, an area where Claude models have been widely watched.

Pricing and model access

Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, the same listed pricing as Claude Sonnet 5. Input tokens are what users send to the model, while output tokens are what the model generates in response.

Moonshot AI said the full model weights will be released by July 27 under a modified MIT license. Model weights are the files that let developers run, fine-tune or build services on top of an AI model locally, rather than only accessing it through a company-hosted chatbot or application programming interface.

The company also highlighted two technical features in its Kimi K3 announcement. Moonshot AI said Kimi Delta Attention enables up to 6.3 times faster decoding in million-token contexts, referring to the process of generating responses. It also said Attention Residuals deliver about 25% higher training efficiency at less than 2% additional overhead.

The release adds another major Chinese model to the global AI competition, at a time when U.S. labs such as Anthropic and OpenAI continue to dominate much of the commercial conversation. Kimi K3’s open-weight approach gives developers a different path: waiting for the model files, then testing whether the benchmark gains translate into real products.

This story draws on original reporting from Decrypt.