Kimi K3

Short answer. Kimi K3 is Moonshot AI's flagship model, unveiled on 16 July 2026. Its mixture-of-experts architecture totals 2.8 trillion parameters, of which 50 billion are active per token, with a one-million-token context window, native vision and always-on reasoning. It ranks third on the Artificial Analysis index at launch, at the level of closed frontier models, ahead of Claude Opus 4.8, and first on the Arena Frontend Code leaderboard. Moonshot announces open weights for 27 July 2026 under a modified MIT license, which would make it the largest open model published.

What Kimi K3 brings

Kimi K3 targets enterprise AI: general-purpose workloads, large-scale agents and long documents. The architecture spreads 2.8 trillion parameters across 896 experts, of which 16 activate per token. Each forward pass therefore mobilizes only about 1.8 percent of the parameters, containing compute cost per request. Kimi Delta hybrid attention speeds up processing of long context, up to one million tokens.

Unlike the K2 generation, reasoning now runs continuously: Moonshot calls it "reflection mode." The model natively understands images. Two variants exist at launch: K3 Max for conversation and agentic tasks, K3 Swarm Max for massive parallel processing. Moonshot announces a training efficiency 2.5 times that of K2.

Technical specifications

Characteristic	Value
Total parameters	2.8 trillion (mixture of experts)
Active parameters per token	50 billion (16 of 896 experts)
Context	1 million tokens
Modalities	Text and native vision
Reasoning	Always on ("reflection mode")
Attention	Kimi Delta Attention (linear hybrid)
Quantization	Quantization-aware training: MXFP4 weights, MXFP8 activations
Artificial Analysis index	Around 57 at launch, top open model
Open weights	Announced for 27 July 2026, modified MIT license
API pricing	3 dollars per million input tokens, 15 for output, 0.30 for cache

Positioning against other open models

At its unveiling, Kimi K3 ranks third on the Artificial Analysis index, behind two proprietary American models, and ahead of every open model. It surpasses GLM 5.2 on this index, until then the platform's top open model. Its API pricing places it at the level of mid-range American offerings, above DeepSeek V4 and GLM 5.2: on-premises deployment with free KV cache further strengthens the economic case for self-hosting agentic workloads. Kimi models already power production products, including the Cursor code editor and DoorDash operations.

Kimi K2.7 Code stays in the catalog for coding assistants: lighter, it keeps tool calls stable across long sessions. K3 complements it for general-purpose, multimodal and reasoning workloads.

At the level of frontier models

The Artificial Analysis index credits Kimi K3 with roughly 57 points at launch: three points behind Claude Fable 5, two behind GPT 5.6, one ahead of Claude Opus 4.8. A model announced with open weights thus measures up against leading closed systems, with only a few weeks' gap in release timing. Moonshot also announces a cost per task roughly half that of Claude Opus 4.8.

Two results stand out from the benchmarks. On visual quality, Kimi K3 takes first place on the Arena Frontend Code leaderboard, a blind ranking where developers compare generated interfaces, ahead of Claude Fable 5. On endurance, it finishes first on four of eight real-task automation benchmarks, including SpreadsheetBench 2 and BrowseComp, and scores 90.4 on a one-million-token evaluation with no context management. These long workloads, where an agent chains hundreds of steps without losing track, match the enterprise use cases the model targets.

A methodological caveat applies: these figures come from Moonshot or from API-based measurements. The community can only reproduce them once the weights are published.

Fewer refusals for research and regulated professions

Early usage reports point to behavior less restrictive than closed APIs on professional domains: a biology researcher, a lawyer or a physician gets substantive answers where other models refuse the question or redirect it to a reduced model. For research teams and regulated professions, the model then serves as a genuine working interlocutor, within applicable law and each profession's code of ethics.

Open weights extend this freedom: the operator audits the model's behavior, adjusts it through further training and sets its own application-level guardrails, at the LiteLLM gateway or the agent harness level. Usage policy belongs to the organization that deploys, under its own responsibility, not to the model's publisher.

Running it on-premises

The full model requires a cluster of H200, B200 or B300 servers, or a GB300 NVL72 rack: the 2.8 trillion parameters must reside in memory, even though only 50 billion activate per token. Quantization-aware training, carried out in MXFP4 from supervised fine-tuning onward, prepares the model to serve at 4 bits with no calibration step on the operator's side. Moonshot recommends nodes of 64 accelerators and above, to keep inter-expert traffic within a single high-bandwidth domain. The vLLM engine serves this type of mixture-of-experts architecture, and the LiteLLM gateway makes it interchangeable with the platform's other models: switching models comes down to changing one configuration line.

The release of the weights will also open the way to distilled versions: smaller models, trained on K3's outputs, that bring part of these capabilities within reach of a DGX Spark workstation or similar. The first are expected in the weeks following the weights going live.

One caveat holds until actual publication: as long as the weights and license are not live, on-premises deployment remains a vendor promise. QDNA checks the license and commercial usage terms before any production rollout. The economic comparison between on-premises deployment and API use appears in our study on LLM costs.

Frequently asked questions

What is Kimi K3?

Moonshot AI's flagship model, unveiled on 16 July 2026: 2.8 trillion parameters in a mixture of experts, 50 billion active, a one-million-token context window, native vision and always-on reasoning. Third on the Artificial Analysis index at launch.

Does Kimi K3 reach the level of closed models?

The Artificial Analysis index credits it with roughly 57 points at launch: three behind Claude Fable 5, one ahead of Claude Opus 4.8. It also takes first place on the Arena Frontend Code leaderboard, a blind ranking of interface code, ahead of Claude Fable 5. These figures come from Moonshot or from API-based measurements and remain to be confirmed once the weights are published.

Is Kimi K3 an open-weight model?

Not yet at the time of its unveiling. Moonshot announces the release of the weights for 27 July 2026 under a modified MIT license. Until then, the model remains accessible via Moonshot's API.

What hardware is needed to run it on-premises?

An H200, B200 or B300 cluster, or a GB300 NVL72 rack: the full set of weights must reside in memory. Quantization reduces the footprint, and the 50-billion-parameter activation limit contains compute per request.

What is the difference from Kimi K2.7 Code?

K2.7 Code, roughly one trillion parameters, serves coding assistants and tool calls. K3 triples the size, adds native vision and always-on reasoning, and targets general-purpose enterprise AI and agents.