Moonshot AI confirmed that full model weights for Kimi K3 will be released by July 27, 2026, according to IBTimes UK. The model features 2.8 trillion parameters and a one million token context window, making it the largest open-weight frontier model to come out of China.
The Financial Times reported, citing anonymous sources, that Kimi K3 is expected to match or surpass Anthropic’s Opus 4.8 in performance. TechCrunch confirmed the FT reporting and noted Moonshot is simultaneously raising fresh capital at a $31.5 billion valuation, up from $20 billion when the company closed a $2 billion round in May 2026.
Architecture and Pricing
Kimi K3 is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, alongside a Stable LatentMoE framework that activates 16 of 896 available experts per task. Moonshot AI claims these architectural changes deliver roughly 2.5 times the scaling efficiency of its K2 predecessor, according to IBTimes UK.
The model includes native visual understanding for processing text and images, and supports long-horizon coding tasks with minimal human supervision. Moonshot says K3 can handle large codebases and coordinate terminal tools across extended engineering workflows.
API pricing is set at approximately $12 per million tokens, positioning K3 closer to premium offerings than the discounted pricing typical of Chinese AI models, per IBTimes UK. Moonshot plans to publish a technical report covering the model’s architecture, training process, and evaluation alongside the weight release.
Open-Source Frontier Competition
Kimi K3’s release adds a third pole to the open-weight frontier model landscape. Teams building self-hosted agent deployments currently choose between Meta’s Llama family (US), Mistral (EU), and DeepSeek (China). Kimi K3 at 2.8 trillion parameters would be the largest open-weight model from any of them.
The timing is significant. As TechCrunch noted, enterprise executives are increasingly questioning the value of paying for closed-source models from OpenAI and Anthropic. Palantir CEO Alex Karp recently argued that companies should take cheaper open-source models and train them for proprietary use rather than submit data to closed model providers.
For enterprise teams with data sovereignty requirements or those running autonomous agents on sensitive workloads, a frontier-level Chinese open-weight model with a one million token context window expands the self-hosting option set considerably. Independent benchmarking will need to confirm the performance claims once weights become publicly available on July 27.