Chinese AI startup Moonshot AI unveiled Kimi K3 — the world’s first open 3T-class model with 2.8 trillion parameters, native vision, and a 1 million-token context window. By the project’s own estimates, overall it trailed only the proprietary Claude Fable 5 and GPT 5.6 Sol.
Kimi K3 is built on the new Kimi Delta Attention (KDA) architecture and Attention Residuals (AttnRes). KDA processes long sequences more efficiently, while AttnRes retrieves information selectively rather than equally from every layer. As a result, the model understands context more deeply and preserves meaning even when handling complex texts.
Sparsity is handled by the Stable LatentMoE framework: only 16 of 896 experts are active at once. Together with new data and training settings, this delivered a 2.5x efficiency gain over K2.
According to the developers, Kimi models held the size record among open models for nine of the past 12 months.
The new model is already available on the website, in Kimi Work, Code and API. The maximum reasoning mode is enabled by default, with others to follow. Full weights and the report will be released on July 27.
Benchmark results
On some tests, K3 outperformed Fable 5 and GPT-5.6 Sol, while lagging on others. It leads on 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 scores 88.3: above Fable 5 (84.6) and just 0.5 below Sol (88.8).
On FrontierSWE, K3 scored 81.2 versus 86.6 for Fable 5, and on HLE-Full — 43.5 versus 53.3. Testing methodologies differed: some models were evaluated via Claude Code, others via Codex or the project’s own KimiCode.
Coding and agent tasks
K3 can run long engineering sessions with minimal human input, navigate large repositories, and control terminal tools.
In a GPU optimization test, model instances worked independently for up to 24 hours on four tasks — including AttnRes, KDA and MLA kernels with a head dimension of 512 — on an NVIDIA H200 and a GPU from another manufacturer. K3 performed on par with Fable 5 and significantly outpaced Opus 4.8, GPT-5.6 Sol and GPT-5.5. In later development stages, an early version of K3 carried out most of this optimization autonomously within the Moonshot team.
Separately, the model wrote MiniTriton from scratch — a compact compiler for GPU kernels with its own IR layer on top of MLIR and generation of PTX code.
Chip design and game development
As a demonstration, K3 independently designed a neural network chip on its own architecture.
The autonomous run took 48 hours: the model used open-source EDA tools and the Nangate 45 nm library. The chip fits in 4 mm², runs at 100 MHz and outputs more than 8700 tokens per second in simulation. 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, code and vision to create game and interactive prototypes from concepts, images and video.
Knowledge work and video
In one case, K3 reproduced the universal I-Love-Q relations from computational astrophysics. It took about two hours instead of one to two weeks for an experienced researcher. The model reviewed more than 20 scientific papers, evaluated over 300 equations of state, found inconsistencies in published formulas, and wrote more than 3,000 lines of Python code, presenting the result in an interactive HTML dashboard.
In another case, K3 prepared an interactive report on 42 years of ASIC chip industry history over 120 rounds of recursive self-improvement, processing more than 11,000 pages from 87 quarterly reports and 99 original PDFs.
Thanks to native video support, K3 can edit footage: in one example, the model created an explanatory animation of its own architecture in the style of 3Blue1Brown; in another, it independently cut a launch teaser from 56 source clips, including shot selection, music synchronization and audio processing. According to Moonshot, such work would take an experienced editor one to two days, and a beginner three to five.
Limitations
The team emphasized that K3 is sensitive to the loss of reasoning history when switching the agent environment mid-session and may show excessive initiative in ambiguous situations. In terms of usability, the model still noticeably lags behind Fable 5 and GPT-5.6 Sol.
In July 2025, Moonshot AI released Kimi K2 — the first model in the lineup with 1 trillion parameters, which quickly became one of the leading open systems for agent tasks.
