Moonshot AI Launches Largest Open AI Model Kimi K3

The Chinese AI startup Moonshot AI has unveiled Kimi K3, the world's first open 3T model featuring 2.8 trillion parameters, native vision, and a context length of 1 million tokens. According to the project's estimates, it ranks just behind proprietary models Claude Fable 5 and GPT 5.6 Sol.

Kimi K3 is built on a new architecture called Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA enables more efficient processing of long data sequences, while AttnRes selectively extracts relevant information from different layers. This results in a deeper contextual understanding and the preservation of meaning even when processing complex texts.

The model's sparsity is managed by the Stable LatentMoE framework, where only 16 out of 896 experts operate simultaneously. Combined with new data and training settings, this has led to a two-and-a-half-fold increase in efficiency compared to K2.

According to the developers, Kimi has held the record for size among open models for nine of the last twelve months.

Source: Kimi Blog.

The new Moonshot AI model is now available on the website, in Kimi Work, Code, and API. The maximum reasoning mode is enabled by default, with additional modes to be introduced later. Full weights and a report will be released on July 27.

Benchmark Results

In some tests, K3 outperformed Fable 5 and GPT-5.6 Sol, but lagged behind in others. It leads in SWE Marathon (42 vs. 35 and 39), BrowseComp (91.2 vs. 88 and 90.4), and Program Bench (77.8 vs. 76.8 and 77.6). On Terminal Bench 2.1, K3 scored 88.3, surpassing Fable 5 (84.6) and falling just 0.5 points short of Sol (88.8).

However, on FrontierSWE, K3 scored 81.2 compared to Fable 5's 86.6, and on HLE-Full, it scored 43.5 against 53.3. The testing methodologies varied: some models were evaluated using Claude Code, while others used Codex or the proprietary KimiCode.

Coding and Agent Tasks

K3 can conduct lengthy engineering sessions with minimal human involvement, navigate large repositories, and manage terminal tools.

In a GPU core optimization test, the model independently worked for up to 24 hours on four tasks—including AttnRes, KDA, and MLA cores with a head size of 512—on NVIDIA H200 and other GPUs. K3's performance matched that of Fable 5 and significantly outperformed Opus 4.8, GPT-5.6 Sol, and GPT-5.5. In later development stages, an early version of K3 autonomously handled much of this optimization within the Moonshot team.

Source: Kimi Blog.

Additionally, the model independently developed MiniTriton—a compact compiler for GPU cores with its own IR layer on top of MLIR and PTX code generation.

Chip Design and Game Development

As a demonstration, K3 autonomously designed a chip for a neural network based on its own architecture.

The autonomous launch took 48 hours: the model utilized open-source EDA tools and the Nangate 45 nm library. The chip fits within 4 mm², operates at 100 MHz, and simulates over 8700 tokens per second. It includes 1.46 million standard cells, 0.277 MB of SRAM, and an INT4 MAC array with built-in dequantization.

K3 also combines 3D reasoning, coding, and vision to create game and interactive prototypes from concepts, images, and videos.

Knowledge Work and Video

In one case, K3 reproduced universal relationships I-Love-Q from computational astrophysics, completing the task in about two hours instead of the one to two weeks typically required by an experienced researcher. The model reviewed over 20 scientific papers, evaluated more than 300 state equations, identified inconsistencies in published formulas, and wrote over 3000 lines of Python code, presenting the results in an interactive HTML dashboard.

In another instance, K3 prepared an interactive report on 42 years of the ASIC chip industry through 120 rounds of recursive self-improvement, processing over 11,000 pages from 87 quarterly reports and 99 original PDFs.

Source: Kimi Blog.

Thanks to its native video capabilities, K3 can edit videos: in one example, the model created an explanatory animation of its architecture in the style of 3Blue1Brown, and in another, it independently edited a teaser for its launch from 56 source clips, including shot selection, music synchronization, and sound processing. According to Moonshot, such work would take an experienced editor one to two days, and a novice three to five.

Limitations

The project team emphasized that K3 is sensitive to losing the history of reasoning when switching agent environments mid-session and may exhibit excessive initiative in ambiguous situations. In terms of usability, the model currently lags behind Fable 5 and GPT-5.6 Sol.

It is worth noting that in July 2025, Moonshot AI released Kimi K2—the first model in the series with 1 trillion parameters, which quickly became one of the leading open systems for agent tasks.