Artificial Intelligence

Kimi K3 Launches—Moonshot’s 2.8T Open Model Challenges Labs

Moonshot AI released Kimi K3 on July 14, 2026: a 2.8T open model with a 1M-token context that challenges the proprietary labs and ships weights by July 27.

Moonshot AI released Kimi K3 on July 14, 2026, a 2.8-trillion-parameter open model that the Beijing lab calls the world’s first openly available 3T-class system.

The model pairs a 1-million-token context window with native vision and a new attention architecture, and Moonshot says it is already live on Kimi.com, Kimi Work, Kimi Code and the Kimi API. The full open weights are promised by July 27.

K3 follows Kimi K2.6, which Moonshot shipped in April 2026, and extends a run in which the company says Kimi models set the upper bound for open-model size in nine of the past twelve months, according to the Kimi blog.

What Kimi K3 Changes in the Open-Model Race

Kimi K3 is built on Kimi Delta Attention and Attention Residuals, two architectural changes Moonshot says improve information flow across long sequences and deep networks. The lab also scaled up Mixture-of-Experts sparsity, activating 16 of 896 experts under a Stable LatentMoE framework.

Moonshot claims the redesign yields roughly 2.5× better scaling efficiency than Kimi K2, letting the model convert compute into capability more cheaply. “Kimi K3 is the first open model to reach 2.8 trillion parameters,” the company wrote in its launch post, a claim echoed on its main site at Moonshot.

The open-weight field is moving fast on both sides of the Pacific: Thinking Machines recently shipped a 975B open-weight model of its own. On a GPU-kernel optimization benchmark, Moonshot says K3 performed competitively with Anthropic’s Claude Fable 5 and clearly beat OpenAI’s Opus 4.8, GPT 5.6 Sol and GPT 5.5, while conceding the top proprietary systems still lead overall.

How Developers and Rivals Are Reacting to Kimi K3

Within days of launch, developers began wiring K3 into agent workflows. A Hacker News poster published a “Show HN” describing Hermes agents running on Kimi K3 inside a sandbox, and a separate thread compared K3 against Claude Fable 5 and GPT 5.6 Sol in a video benchmark.

The coding angle matters because Google recently delayed Gemini 3.5 Pro after its coding fell short of internal bars, a sign that long-horizon engineering is the front line of model competition. Moonshot is pitching K3 squarely at that workload, citing kernel and GPU-compiler results achieved on NVIDIA H200 hardware.

According to the launch post and company background on Wikipedia, Moonshot will publish the complete K3 model weights by July 27, 2026, after coordinating the rollout with inference partners and open-source maintainers.

FAQ

What is Kimi K3?

Kimi K3 is Moonshot AI’s 2.8-trillion-parameter open model, released July 14, 2026, with native vision, a 1-million-token context, and a Mixture-of-Experts design that activates 16 of 896 experts.

Why does Kimi K3 matter?

It is the first openly released 3T-class model, extending China’s push at the scaling frontier and giving developers a large open alternative to proprietary systems from Anthropic and OpenAI.

What comes next for Kimi K3?

Moonshot plans to publish the full model weights by July 27, 2026, and to add low- and high-effort thinking modes after launch, while coordinating the rollout with inference partners.