Moonshot AI, the Beijing-based lab behind the Kimi chatbot, has launched Kimi K3, a 2.8-trillion-parameter multimodal model, and made it available immediately across its consumer app, desktop client, coding tool, and API. Full open weights are scheduled for July 27. At 2.8 trillion parameters, Kimi K3 is now the largest open-weight model on the market, ahead of DeepSeek’s V3 family and Alibaba’s Qwen3, the two labs that have anchored the open-weight tier since 2025.
The model carries a 1-million-token context window and native vision input, meaning it can read screenshots, video frames, and long documents in the same session without a separate multimodal adapter. Moonshot built K3 on two new architectural components it calls Kimi Delta Attention and Attention Residuals, paired with a sparser mixture-of-experts setup that activates 16 of 896 experts per token. The company says this combination delivers roughly 2.5 times the scaling efficiency of its prior flagship, Kimi K2, meaning more capability per unit of training compute rather than a straightforward parameter increase.
Moonshot positions K3 as a coding-first agent, citing case studies where the model optimized GPU kernels, built a Triton-like compiler called MiniTriton from scratch, and completed an astrophysics research pipeline (reviewing more than 20 papers and writing over 3,000 lines of Python) in about two hours. In its own kernel-optimization tests, the company reports K3 outperformed Claude Opus 4.8 and both current GPT 5 variants, and ran competitively with Claude Fable 5. Every one of those comparisons comes from Moonshot’s internal evaluation suite, not an independent benchmark operator, and the company acknowledges elsewhere in its release notes that K3’s overall capability still trails Claude Fable 5 and GPT 5.6 Sol, the two proprietary models it is chasing.
Pricing on the Kimi API is $0.30 per million tokens for cached input, $3.00 per million tokens for uncached input, and $15.00 per million tokens for output, undercutting the token rates that Anthropic and OpenAI charge for their top-tier models. Moonshot also disclosed two operational caveats worth flagging for any team planning to deploy K3: the model is sensitive to how its reasoning history is passed back between turns and can degrade if a harness drops that context, and it was trained to act with unusual autonomy, which the company warns can lead it to make unrequested decisions unless a system prompt constrains it explicitly. Moonshot says K3 still lags Claude Fable 5 and GPT 5.6 Sol on general user experience, a rare admission of a gap it has not closed.
The scaling race among open-weight labs has moved fast enough that “largest” is now a short-lived title: DeepSeek and Qwen have each held it within the past year, and Moonshot itself says Kimi models have set the open-model size ceiling for nine of the past twelve months. What separates K3 from a pure size claim is the pairing of frontier-scale parameters with a 1-million-token window and a per-token price closer to a mid-tier API than a flagship one, a combination no other open lab currently offers at this scale.
Teams evaluating open-weight alternatives to Claude or GPT for long-context or agentic coding workloads should hold off on production decisions until the July 27 weights drop allows independent benchmarking outside Moonshot’s own test suite.
This account is based on Moonshot AI’s official Kimi blog post announcing Kimi K3.
