The race to build the biggest open-source AI model just took another leap forward. Chinese startup Moonshot AI has released the full open weights of Kimi K3, handing developers free access to what is now the largest open-weight AI model ever made publicly available. At 2.8 trillion parameters, Kimi K3 pushes open models into territory that was once reserved for a handful of well-funded AI labs, adding fresh momentum to the growing competition between open and proprietary AI systems.

The release became available at 00:00 UTC on July 27 through Hugging Face, where developers can download the complete model weights. Stored using MXFP4 quantization, the download occupies roughly 1.4 terabytes, making it one of the largest AI model releases ever distributed to the public.

In a post on X announcing the release, Moonshot AI wrote:

“Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params.”

The company added, “Alongside Kimi K3, we’re opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.”

The release comes just three days after OpenAI and Google joined a growing coalition backing open-weight AI, with Elon Musk endorsing Nvidia CEO Jensen Huang’s call for American AI leadership.

Running the model locally remains a challenge. Moonshot says self-hosting typically requires hardware such as eight Nvidia H100 GPUs or more than 2TB of VRAM, placing it beyond the reach of most individual developers. Still, access to the weights opens the door for research labs, cloud providers, enterprises, and well-funded startups that want to fine-tune or deploy the model without relying on a commercial API.

Kimi K3 Launches With 2.8 Trillion Parameters, Overtakes Claude Fable 5 in Frontend Coding

The milestone arrives just eleven days after Moonshot AI introduced Kimi K3, its flagship large language model. Within hours of its debut, Kimi K3 climbed to the top of Arena.ai’s Frontend Code Arena, a benchmark that measures how well AI systems build complete web applications from natural language prompts. The ranking placed Kimi K3 ahead of Anthropic’s Claude Fable 5, giving the Chinese startup an early victory in one of AI’s most closely watched coding benchmarks.

At the heart of Kimi K3 is a Mixture-of-Experts architecture containing roughly 2.8 trillion parameters. Rather than activating the entire model for every request, Kimi K3 activates only 16 of its 896 experts for each token. That approach reduces computing requirements during inference without sacrificing much of the model’s capability. Moonshot pairs that architecture with a context window of up to one million tokens, allowing the model to process entire software repositories, lengthy technical documentation, legal records, research papers, and other large collections of information in a single session.

Moonshot released two versions of the model. K3 Max focuses on general conversation, reasoning, coding, and autonomous agent tasks. K3 Swarm Max targets multi-agent orchestration, allowing several AI agents to collaborate on larger software engineering and research projects. The split reflects a broader shift across the AI industry as companies move beyond chatbots and invest in AI systems that can complete extended workflows with limited human supervision.

Early benchmark results suggest Kimi K3 is more than a model built around headline parameter counts. According to Artificial Analysis, Kimi K3 ranks second overall on AA-Briefcase, the firm’s benchmark for agentic knowledge work. It trails only Claude Fable 5 and finishes ahead of models including GPT-5.6 Sol, Claude Sonnet 5, and Claude Opus 4.8.

Artificial Analysis reports that Kimi K3 earned an AA-Briefcase Elo score of 1,543, a dramatic improvement over the previous Kimi K2.6 generation. The model posted a 51% rubric pass rate, placing it just behind Claude Fable 5’s 56%. Its analytical reasoning scores rank among the strongest measured so far, though presentation quality still lags behind several competing models.

Credit: Artificial Analysis

Those gains come with meaningful tradeoffs. Artificial Analysis estimates that Kimi K3 averages $10.57 per benchmark task, roughly ten times the cost of Kimi K2.6. The model completes an average task in about 56 minutes, driven by long conversations, larger outputs, and slower inference speeds through Moonshot’s first-party API. That places it among the most expensive frontier models to operate despite its impressive reasoning performance.

The release carries significance beyond benchmark rankings. Open-weight models give developers the ability to inspect, modify, fine-tune, and deploy AI systems on their own infrastructure without depending entirely on API providers. Enterprises can customize models for proprietary datasets, research institutions can study model behavior in greater detail, and startups gain another high-performance alternative to commercial closed systems.

Moonshot’s decision lands at a time when the debate over open versus proprietary AI continues to intensify. U.S. companies such as OpenAI and Anthropic still keep their frontier models closed, arguing that unrestricted releases increase security and misuse risks. Chinese AI firms have increasingly taken a different path, publishing open-weight models that encourage broader adoption across the research community and enterprise developers.

The latest release adds another chapter to that competition. Over the past year, Chinese companies including DeepSeek, Alibaba, Z.ai, MiniMax, and Moonshot AI have steadily narrowed the performance gap with leading U.S. models across reasoning, coding, and agentic workloads. Kimi K3’s release signals that the contest is no longer centered solely on benchmark scores. It now includes who can place the most capable AI systems directly into developers’ hands.

For software teams, researchers, and companies building AI products, free access to a 2.8 trillion-parameter model offers something that was almost unthinkable a year ago. The largest open-weight AI model ever released is no longer confined to one company’s servers. It is available for anyone with enough computing infrastructure to put it to work.