Moonshot drops Kimi K3 weights for free, beats Fable 5 in key benchmarks

Moonshot AI has made its flagship Kimi K3 model free for developers to download, customise and run. The company claims the model beats Anthropic's Fable 5 on several coding benchmarks while also open-sourcing the core technologies that power its AI.Moonshot just launched Kimi K3, and it can almost match the performance level of Claude Fable or GPT-5.6 (Representational image made with AI).Moonshot AI has released the model weights and technical report for Kimi K3, calling it its most capable AI model yet. The release marks one of the biggest open-source AI releases of the year. The Beijing-based company is now making the model freely available to developers, along with the core technologies used to train and run it. This means researchers and businesses can download, modify and host Kimi K3 on their own systems.According to Moonshot AI, Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts (MoE) model with native visual understanding and a one-million-token context window. The company says the release is aimed at speeding up the adoption of advanced AI while supporting research towards Artificial General Intelligence (AGI). Alongside the model weights, Moonshot has also open-sourced FlashKDA, MoonEP and AgentEnv, three core technologies used to train and run Kimi K3.“Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window,” the company said in its official announcement on X. “Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.” But why is this a big deal? Unlike companies such as OpenAI and Anthropic, which keep their flagship models closed, Moonshot is giving developers the actual model weights for free. This will allow developers to study the model, customise it, and deploy it on their own infrastructure instead of depending on Moonshot's API. At around 1.4TB, Kimi K3 is also among the largest open-weight AI models released so far. Founder Yang Zhilin has previously said openness is central to the company's strategy for attracting developers and businesses. Moonshot AI says Kimi K3 is nearly three times larger than Kimi K2.5. But the company says the real improvement comes from a new architecture rather than simply adding more parameters. It claims technologies such as Kimi Delta Attention, Attention Residuals and MoonEP improve scaling efficiency by 2.5 times compared to the previous generation. The company has also published a detailed technical report explaining how the model was trained.The technical report also explains some of the improvements behind Kimi K3. Moonshot says the new model is better at handling long conversations, understanding images and completing complex reasoning tasks. It also introduces a new training method that combines different AI skills into a single model, making Kimi K3 more capable across coding, maths, reasoning and other tasks.Moonshot AI has also open-sourced the technologies that power Kimi K3. The company says these tools help the model run faster, train more efficiently and support large-scale AI workloads. By making them freely available, Moonshot hopes developers can build and deploy AI applications more easily.Kimi K3 beats Fable 5 in some benchmarksMoonshot says Kimi K3 performs strongly across coding, reasoning, mathematics and agentic tasks. According to the benchmark charts published in its technical report, the model outperforms Anthropic's Fable 5 on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench and SWE-Marathon. However, the company also notes that Claude Fable 5 still ranks ahead of Kimi K3 in its overall evaluation.- EndsPublished By: Divya BhatiPublished On: Jul 28, 2026 11:07 IST