Moonshot Unveils 2.8-Trillion-Parameter Kimi K3 Model

Moonshot's documentation attributes K3's architecture to Kimi Delta Attention (KDA), described as a hybrid linear-attention mechanism, and Attention Residuals (AttnRes). According to the documentation, the system uses the Stable LatentMoE framework and activates 16 of 896 experts, a sparse mixture-of-experts configuration intended to limit active computation relative to the model's total parameter count.

Moonshot claims these architectural and training changes provide roughly 2.5 times K2's scaling efficiency. The company also describes K3 as capable of sustaining long-running engineering work with limited human supervision, operating on large codebases, coordinating terminal tools, and combining software tasks with visual inputs such as screenshots. Those are vendor claims, and the company stated that further architecture, training, and evaluation details would accompany a technical report.

For ML infrastructure teams, the 2.8-trillion total parameter figure should not be read as a direct measure of per-request compute. In sparse MoE systems, the number of activated experts, expert routing, sequence length, precision format, and serving stack are more immediately relevant to latency and memory planning. A 1 million-token context window also creates separate questions around usable long-context quality, attention-memory behavior, retrieval design, and cost under production workloads.

Reuters reported that Moonshot said K3 performed competitively with Anthropic's Fable 5 with fallback on GPU-kernel optimization and outperformed Anthropic Opus 4.8, OpenAI GPT 5.6 Sol, and GPT 5.5 in that evaluation. CNBC separately reported that Moonshot said K3 trailed Fable 5 and GPT 5.6 Sol on overall performance while outperforming other tested systems on benchmarks that included coding and general-agent tasks.

Reuters also reported that Arena.ai ranked K3 first in a web interface-building benchmark and that Vals AI placed it second overall behind Fable 5. Such rankings can be useful directional evidence, but practitioners typically need task-level data, test-set provenance, prompting details, tool configurations, and reproducible serving conditions before translating benchmark positions into production-model selection.

Bank of America analysts, quoted by CNBC, said that K3 demonstrated how pre-training scale and architectural innovation can deliver substantial gains despite China's hardware and compute constraints. More broadly, comparable open-weight releases expand the set of models that organizations can inspect and customize, while shifting evaluation work toward practical concerns such as weight access, hardware compatibility, inference throughput, safety controls, and benchmark reproducibility.

Moonshot's documentation states that it is working with inference partners and open-source maintainers to align technical details ahead of the weight release. Until the weights and technical report are available, independent assessment of K3's reproducibility and real-world serving characteristics remains limited.