Kimi K3

Short answer. Kimi K3 is Moonshot AI's flagship model, unveiled in July 2026. Its mixture-of-experts architecture totals 2.8 trillion parameters, of which 104 billion are active per token, with a one-million-token context window, native vision and always-on reasoning. As of 2 September 2026, the Artificial Analysis index credits it with 60 points in max mode, tied with GLM 5.3 at the top of open-weight models, behind Claude Fable 5.1 (66) and Claude Opus 5 (63). The weights are published on Hugging Face under the model-specific Kimi K3 licence: 1,561 GB in MXFP4, the largest open model published according to Moonshot.

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% 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. Moonshot announces a scaling efficiency (conversion of compute into capability) roughly 2.5 times that of K2, and max thinking effort by default at launch, with low- and high-effort modes announced for later updates.

Technical specifications

Characteristic	Value
Total parameters	2.8 trillion (mixture of experts)
Active parameters per token	104 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	60 points in max mode as of 2 September 2026, tied with GLM 5.3 at the top of open-weight models
Open weights	Published on Hugging Face, Kimi K3 licence: 1,561 GB in MXFP4, 93 layers (config.json and repository re-read on 2 September 2026)
API pricing	3 dollars per million input tokens, 15 for output, 0.30 for cache

Positioning against other open models

As of 2 September 2026, Kimi K3 shares the top of open-weight models on the Artificial Analysis index with GLM 5.3, at 60 points, ahead of GLM 5.2, 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 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 60 points in max mode as of 2 September 2026, against 66 for Claude Fable 5.1 and 63 for Claude Opus 5. An open-weight model thus measures up against leading closed systems.

On endurance, Moonshot publishes its results on real-task automation benchmarks (SpreadsheetBench 2, OfficeQA Pro, BrowseComp, Terminal-Bench 2.1) and a score of 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, and the publication of the weights now allows independent reproduction.

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 104 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 also opens 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.

One caveat remains on the licence: the Kimi K3 License is specific to this model and does not carry over the modified MIT of Kimi K2.7 Code, so its commercial usage terms must be read before any deployment. QDNA checks the licence 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.

Official sources

• moonshotai/Kimi-K3, official model card: licence, parameters, context window and weight format
• Artificial Analysis, model comparison index

Frequently asked questions

What is Kimi K3?

Moonshot AI's flagship model, unveiled in July 2026: 2.8 trillion parameters in a mixture of experts, 104 billion active, a one-million-token context window, native vision and always-on reasoning. Top of open-weight models on the Artificial Analysis index as of 2 September 2026, tied with GLM 5.3.

Does Kimi K3 reach the level of closed models?

The Artificial Analysis index credits it with 60 points in max mode as of 2 September 2026, against 66 for Claude Fable 5.1 and 63 for Claude Opus 5. The other figures come from Moonshot or from API-based measurements and can now be reproduced since the weights were published.

Is Kimi K3 an open-weight model?

Yes. Moonshot publishes the weights on Hugging Face under the Kimi K3 licence, a licence specific to this model rather than the modified MIT of Kimi K2.7 Code. The model also remains accessible through 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 104-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.