Kimi K3

Moonshot AI's most capable model: an open-weight, natively multimodal agentic model with 2.8T total parameters and roughly 104B activated across 896 experts. It is built on Kimi Delta Attention and Attention Residuals with a Stable LatentMoE framework, giving roughly 2.5 times better scaling efficiency than Kimi K2, and pairs native visual understanding with a million-token context. It is designed for frontier work -- software engineering, knowledge work and deep reasoning -- and always reasons.

Moonshot's own framing is that the parameter count is not the point. The architecture changes are what the lab argues for: Kimi Delta Attention as an efficient base for scaling attention, and Attention Residuals, which retrieve representations selectively across depth instead of accumulating them uniformly. Sparsity was pushed further too, with a Stable LatentMoE framework effectively activating 16 of 896 experts, and the routing work that makes that stable at scale - allocation derived directly from router-score quantiles rather than from a hand-tuned balancing term, and an optimiser that treats each attention head independently. The claim Moonshot puts on all of it together is roughly 2.5 times better scaling efficiency than Kimi K2: compute converted into capability, rather than parameters converted into a headline.

The case for the model is made mostly in case studies. Given 24 hours per task in identical sandboxes, Moonshot reports K3 competitive with Claude Fable 5 and substantially ahead of Claude Opus 4.8, GPT 5.6 Sol and GPT 5.5 at optimising GPU kernels across two hardware families, and notes that late in K3's own development an early version handled most of the team's kernel work. Asked to build a GPU programming system from scratch it produced a compact Triton-like compiler with its own tile-level intermediate representation, optimisation passes and a code-generation path down to PTX, which the lab says matches or beats the established stack on supported benchmarks and trains a small transformer end to end. In a single 48-hour autonomous run it designed, optimised and verified a chip for a nano model built on its own architecture, closing timing at 100 MHz inside four square millimetres.

The published table backs that with 88.3 on Terminal-Bench 2.1, 42.0 on SWE Marathon and 77.8 on Program Bench, the highest in its compared set on the last two, alongside 93.5 on GPQA-Diamond and 91.2 on BrowseComp. Multimodality is native rather than bolted on, and Moonshot leans on it for video in particular: it points to K3 cutting its own launch teaser from 56 source clips with beat-synchronised edits, and producing an animated explainer of its own architecture.

The limitations section is the part to read before wiring it into anything. K3 was trained in preserved-thinking mode, so a harness that fails to pass the full thinking history back, or a session handed over to K3 from another model mid-way, can make generation highly unstable. It is also, in the lab's own words, prone to excessive proactiveness: trained hard on long-horizon work, it tends to decide on your behalf when the intent is ambiguous, and needs explicit behavioural constraints in the system prompt if it must stay inside defined boundaries. Moonshot states plainly that overall performance still trails Claude Fable 5 and GPT 5.6 Sol, with a noticeable gap in user experience.

The weights are open. Moonshot committed to publishing the full set by 27 July 2026 and contributed a prefix-caching implementation for its new attention design upstream so the model can be served efficiently outside the lab's own stack. Self-hosting it is a serious undertaking rather than a download, though: the recommendation is deployment on supernode configurations of 64 or more accelerators. The copy served here is Moonshot's.