Moonshot Releases Kimi K3 Weights and Report

Moonshot's documentation identifies two core architectural components: Kimi Delta Attention (KDA), described as a hybrid linear-attention mechanism, and Attention Residuals (AttnRes), which changes the residual-path design across model depth. The company states that these architectural changes, combined with training and data-recipe improvements, produce roughly 2.5x the overall scaling efficiency of Kimi K2.

Those details are notable because attention cost and memory traffic become central constraints at million-token context lengths. Sparse MoE reduces active computation per token, while linear or hybrid attention designs target the quadratic scaling cost associated with standard full attention. Claims of end-to-end scaling efficiency, however, depend on training setup, hardware, data, and evaluation methodology, so they are not directly comparable across labs without more complete reproducibility artifacts.

Moonshot describes K3 as its first open model in the 3-trillion-parameter class. Reuters similarly reported that the model is the largest open-weight AI system announced to date, while explaining that open-weight releases let users download, run, and customize the underlying model rather than relying solely on a hosted proprietary endpoint.

Moonshot's blog frames K3 around long-horizon coding, reasoning, and knowledge-work tasks. The company reports that the model can work across large repositories, coordinate terminal tools, and combine visual feedback with software-engineering workflows such as frontend development, game development, and CAD.

Third-party benchmark reporting provides some external context, although benchmark scope matters. Reuters reported that Arena.ai ranked K3 first for web interface-building capability and that Vals AI placed it second overall behind Fable 5. Tom's Hardware reported a 1,679-point Frontend Code Arena result, ahead of Fable 5 in blind developer testing. Moonshot's own evaluation claims should be treated separately: its blog states that K3 trails Claude Fable 5 and GPT 5.6 Sol overall, while outperforming other tested models in its evaluation suite.

For ML teams, the release makes a very large sparse model available for evaluation and self-hosted inference rather than only API consumption. In comparable open-weight releases, the practical constraint is rarely parameter access alone: deployment depends on quantization quality, expert-routing support, kernel maturity, distributed serving infrastructure, and the available accelerator memory and interconnect bandwidth.

VentureBeat reported that K3 supports an OpenAI- and Anthropic-compatible API interface, which could reduce application-level migration work for teams already using those schemas. It reported listed API pricing of $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached-input tokens.

36Kr reported Day 0 support announcements from Nebius, Baseten, Fireworks, Huawei Ascend CANN, and an Ascend 910C deployment adaptation based on SGLang. These announcements indicate early availability across multiple inference stacks, but production users still need to validate throughput, latency, tool-use reliability, multimodal behavior, and license terms against their own workloads.

The K3 release gives practitioners an unusually large open-weight system to benchmark for long-context and agentic workloads. Its real operational significance will depend on whether the surrounding ecosystem can make a 896-expert architecture economical and reliable outside controlled evaluations.