Kimi K3

Kimi K3 is Moonshot AI's July 2026 frontier model for long-horizon coding, agentic knowledge work, and multimodal reasoning. At 2.8 trillion total parameters, it is the largest model announced for an open-weight release. K3 is available through Kimi's apps, Kimi Code, and the Moonshot API, while the full downloadable weights and technical report are promised by July 27, 2026.

Core profile

K3 is a sparse mixture-of-experts model with 896 total experts and 16 activated for each token. It accepts text and images, produces text, and supports a 1 million token context window. Reasoning runs at maximum effort by default at launch, with lower-effort modes planned for later updates.

The model sits in the proprietary frontier tier before its weights are public. Artificial Analysis scored K3 at 57 on its independently run Intelligence Index, placing it third overall at launch behind Claude Fable 5 and GPT-5.6 Sol, and in the same capability tier as Claude Opus 4.8 and GPT-5.5. If Moonshot releases the weights under usable terms, K3 will become the strongest open-weight model Artificial Analysis has measured.

Architecture

K3 combines Kimi Delta Attention, or KDA, with Attention Residuals. KDA is designed to make attention scale more efficiently across long sequences, while Attention Residuals selectively retrieve representations across model depth instead of accumulating every layer uniformly.

Moonshot uses a framework called Stable LatentMoE to keep routing stable at extreme sparsity. Quantile Balancing assigns expert load from router-score quantiles, while Per-Head Muon optimizes attention heads independently. The goal is to prevent a small number of experts from becoming overloaded while others remain underused as the model scales.

K3 also uses quantization-aware training from the supervised fine-tuning stage onward, with MXFP4 weights and MXFP8 activations. Moonshot presents this as a route to broad hardware compatibility, but K3 is not a small local model. The company recommends supernode deployments with 64 or more accelerators, so self-hosting requires data-center-scale infrastructure even after the weights become available.

Benchmark profile

K3's strongest launch results are concentrated in coding and agentic work. The comparison below shows where it leads and where the two strongest proprietary models remain ahead.

Evaluation	Kimi K3	Claude Fable 5	GPT-5.6 Sol
Artificial Analysis Intelligence Index	57	60	59
DeepSWE	67.5%	70.0%	73.0%
Program Bench	77.8%	76.8%	77.6%
Terminal-Bench 2.1	88.3%	84.6%	88.8%
FrontierSWE	81.2%	86.6%	71.3%
SWE Marathon	42.0%	35.0%	39.0%
PostTrain Bench	36.6%	41.4%	34.6%
GDPval-AA v2, Elo	1668	1760	1748
BrowseComp	91.2%	88.0%	90.4%

The Intelligence Index row comes from independent Artificial Analysis testing. The remaining values are compiled in Moonshot's launch table from a mixture of vendor and third-party evaluations. Bold marks the highest reported result in each row, not a universal winner. Fable 5 was tested at maximum effort with fallback behavior where applicable, and the coding rows use different agent harnesses.

The table shows a more useful pattern than a single ranking. K3 leads both comparison models on Program Bench, SWE Marathon, and BrowseComp. It nearly matches GPT-5.6 Sol on Terminal-Bench, sits between Fable 5 and Sol on FrontierSWE and PostTrain Bench, and trails both on DeepSWE and broad composite intelligence. K3 is not the best model across the board; it is already trading wins with the two strongest proprietary systems.

K3's DeepSWE result is 67.5% with Kimi Code and 67.3% with the benchmark's mini-SWE-agent harness. That two-tenths gap is evidence that its score is not entirely created by Moonshot's own wrapper, even though both Fable 5 and Sol remain ahead in the launch comparison.

The Terminal-Bench comparison needs more caution. K3 ran through Kimi Code, Claude models through Terminus 2, and GPT models through Codex. A benchmark score measures the model and the harness together, so results produced through different agent systems are directional rather than perfectly interchangeable.

Arena's Frontend Code Arena result adds a different kind of evidence. Users judged outputs blindly and preferred K3 over every leading American model tested, including Claude Fable 5 and GPT-5.6 Sol, for that specific category of front-end work. That does not make K3 the best model overall; it shows that an open-weight candidate can already lead the proprietary frontier on a meaningful task.

Pricing and efficiency

Moonshot prices K3 at USD 0.30 per million cached input tokens, USD 3 per million cache-miss input tokens, and USD 15 per million output tokens. The company says its Mooncake serving system reaches cache-hit rates above 90% on coding workloads, which can reduce repeated repository-input costs substantially.

The low cached-input rate does not make every K3 task cheap. Artificial Analysis measured an average cost of USD 0.94 per Intelligence Index task, compared with USD 1.80 for Claude Opus 4.8 and USD 1.04 for GPT-5.6 Sol. K3 was roughly half the cost of Opus but only slightly cheaper than Sol, and it remained more expensive than open-weight peers such as GLM 5.2 and DeepSeek V4 Pro.

Output volume explains part of the difference. K3 generated roughly 132 million output tokens across the Artificial Analysis evaluation suite. Always-on maximum reasoning can consume enough tokens to narrow the advantage suggested by the API rate card.

The open-weight caveat

K3 is available as a hosted model today, but its open-weight status remains a commitment until the checkpoint and license are public. Without the weights, outside researchers cannot fully inspect the architecture, reproduce deployment behavior, test quantized checkpoints, or verify that self-hosted K3 matches Moonshot's API.

The July 27 release therefore matters more than another launch benchmark. If the checkpoint arrives on time under terms that permit inspection, fine-tuning, and deployment, K3 becomes both the largest downloadable model and the leading open-weight model in independent composite evaluation. If the release slips or the license restricts meaningful use, the open-weight label will have described a roadmap rather than a shipped artifact.

When to use it

K3 is a strong candidate for repository-scale coding, terminal agents, front-end generation, long-document analysis, visual software tasks, and sustained knowledge work where high capability matters more than minimum output cost. Its cache economics may be attractive for repeated work over the same large codebase or document collection.

When not to use it

K3 is not the obvious choice for lightweight classification, high-volume low-stakes requests, latency-sensitive chat, or local deployment on ordinary workstations. Teams that require auditable open weights today should wait for the checkpoint, license, and technical report rather than treating the promised release as complete.

Bottom line

Kimi K3 is already a frontier model and a serious coding system. What remains unproven is the claim that makes it structurally different from Claude or GPT: that a 2.8-trillion-parameter model this capable is about to become something developers can inspect, fine-tune, and control rather than only rent through an API.