Moonshot AI K3 Launch: A Sleepless Night, Watching Long-Termism Pay Off

To conquer the most difficult and aspire to the most distant goals — launch with courage, persist with focus, achieve with strength.

On the eve of WAIC, BlueRun Ventures' Booming Night wound down close to 11 p.m. The Moonshot AI team, along with many AI founders, talked with us until the very end. Discussions about Agents, open source, world models, and the next frontier of AI's technical boundaries were still going strong.

A few hours later, a new boundary had already emerged.

At 1 a.m. today, Moonshot AI officially released K3 — the world's first open-source model at the 3-trillion scale. The announcement quickly set the tech world abuzz and kept many people up through a "sleepless night." As another major step in pushing the frontier of open-source model scale over the past year, K3 not only elevates China's open-source models to new heights, but demonstrates leading competitiveness in long-horizon programming, knowledge work, reasoning, and other core capabilities.

More noteworthy than the 3 trillion parameters or 1 million token context window is the direction of model evolution that K3 represents. What truly determines long-horizon intelligence is not how much a model can remember, but how the system selects, compresses, extends, and leverages the past. From answering questions to continuously completing complex tasks, model competition is increasingly becoming a contest of real-world task capability and complete engineering systems.

As one of Moonshot AI's earliest investors, BlueRun Ventures was the first to issue a term sheet in the A1 round and has continued to invest across four consecutive rounds. Walking this journey together, we have witnessed the team consistently choosing the hardest paths, repeatedly turning "conquering the most difficult to aspire to the most distant" into reality. Congratulations on the K3 launch, and we look forward to more explorations continuously opening new frontiers of intelligence.

---

Today, we officially launch Kimi K3, our most capable model to date. Kimi K3 is a 2.8 trillion parameter model built on the KDA hybrid linear attention mechanism (Kimi Delta Attention) and Attention Residuals technology, with native visual understanding and a 1 million token context window. It is the world's first open-source model at the 3-trillion scale, designed for frontier intelligence scenarios including long-horizon programming, knowledge work, and reasoning.

While Kimi K3's overall performance still trails the strongest closed-source models Claude Fable 5 and GPT-5.6 Sol, it demonstrates frontier-level capabilities across our full benchmark suite and stably surpasses all other models.

The release of Kimi K3 is just the beginning. We will continue to unlock the potential of the K3 model and steadily improve its performance on real-world tasks.

Starting today, Kimi K3 is available through kimi.com, the latest version of the Kimi mobile app, the latest version of the Kimi Work desktop client, Kimi Code, and the Kimi API. The default thinking intensity is currently set to max (extreme); low and high modes will be added in a future update.

We are currently working closely with inference partners and open-source maintainers to align technical details and ensure reliable deployment across the ecosystem. Full model weights will be released by July 27, 2026. Additional details on architecture, training, and evaluation will be published alongside the Kimi K3 technical report.

A 3-Trillion-Scale Open-Source Model

Kimi K3 is the first open-source model to reach 2.8 trillion parameters. This is the latest step in Moonshot AI's ongoing push to expand model scale boundaries: for 9 of the past 12 months, Kimi models have held the scale ceiling among open-source models.

Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Both architectural updates are designed to let information flow more smoothly through longer sequences and deeper models. We have also further increased the sparsity of the Mixture of Experts (MoE): combined with the Stable LatentMoE framework, the model can efficiently activate 16 out of 896 experts. Together with optimized training methods and data recipes, these structural improvements give Kimi K3 approximately 2.5× better overall scaling efficiency compared to K2, more effectively converting compute into capability.

Programming

Kimi K3 possesses strong long-horizon coding capabilities. With minimal human supervision, it can sustain lengthy engineering tasks, understand and process large codebases, and coordinate terminal tool usage.

Kimi K3 also excels at tasks combining software engineering with visual reasoning. It can leverage screenshots and visual feedback to optimize scenarios such as game development, front-end work, and CAD.

The cases below demonstrate how Kimi K3 translates coding ability into open-ended software development and scientific research capabilities.

Kernel Optimization

We built a kernel optimization arena to test whether Kimi K3 could truly optimize GPU kernels from end to end. Each model ran independently in its own GPU sandbox, with identical tasks, testing frameworks, and production workloads. Given up to 24 hours, the model could analyze existing implementations, rewrite kernels, and iteratively benchmark to validate results.

The arena spans two hardware platforms, three kernel types, and four tasks: Attention Residuals and KDA linear attention on NVIDIA H200 GPUs, a from-scratch 512-head-dimension MLA kernel, and a KDA task on a domestic GPU. At maximum thinking intensity, Kimi K3's performance approaches Fable-5 (with fallback mechanisms) and clearly leads Opus 4.8, GPT-5.6 Sol, and GPT 5.5.

GPU Compiler Development

We further tested whether Kimi K3 could build a GPU programming system from scratch. Kimi K3 developed MiniTriton: a compact Triton-like compiler. Built on MLIR, it constructed its own tile-level intermediate representation and implemented a complete optimization pass to PTX code generation pipeline.

On the Roofline benchmarks it supports, MiniTriton's performance matches or exceeds Triton and torch.compile, and surpasses Triton on some workloads. It currently supports FP32 and FP64 compute; TF32 and BF16 are in development. We did not expect its early Tensor Core implementation to beat highly optimized Triton, but the results show that without benchmark-specific shortcuts or obvious hardcoded optimizations, MiniTriton can already generate competitive GPU code.

As end-to-end validation, we trained nanoGPT with TensorLite and observed normal convergence. This demonstrates that Kimi K3 did more than generate a few isolated kernels — it built a genuinely usable compiler stack: including programming abstractions, intermediate representations, optimization passes, backend code generation, and the ability to connect to real training workloads.

Creating Digital Works

Kimi K3 fuses powerful 3D reasoning, programming, and visual capabilities to transform concepts, images, and videos into fully playable interactive experiences. Seamlessly iterating between code and real-time screenshots, Kimi K3 achieves true vision-in-the-loop — seeing output instantly, optimizing immediately.

Here are some examples:

• 3D simulation of the Long March 10 rocket launch and recovery
• 3D open-world game
• 3D GBA emulator
• Recreating the black hole Gargantua

Visit: https://blackhole-visualizer.ok.kimi.link/

Chip Design

As an early proof of concept, Kimi K3 designed a chip to run a nano model built on its own architecture. During 48 hours of continuous autonomous agent operation, K3 independently completed the chip's construction, optimization, and verification using open-source EDA tools and the Nangate 45nm process library. The chip spans 4 mm², integrates 1.46 million standard cells, 0.277 MB SRAM, and an INT4 MAC array with fused dequantization, achieving timing closure at 100 MHz with simulated decode throughput consistently exceeding 8,700 tokens per second. A chip designed by a model, for a model — this is the portrait of K3's long-horizon agent capability.

(Video generated by Kimi K3 connected to Blender MCP)

Scientific Research Programming

Kimi K3 can bridge scientific literature and executable code, autonomously completing the implementation, verification, and analysis of complex computational research workflows.

In one case, Kimi K3 completed in roughly two hours what would typically take a senior researcher one to two weeks. To reproduce the universal I-Love-Q relations in computational astrophysics, it read and cross-validated over 20 papers, implemented the complete numerical pipeline, evaluated more than 300 equations of state, discovered inconsistencies in published formulas, generated over 3,000 lines of Python code, and produced an interactive HTML dashboard for exploring results.

Knowledge Work Kimi K3 advances end-to-end knowledge work. Beyond public benchmarks, Kimi K3 (max) shows steady improvement in our internal evaluations. These assessments draw from recurring task patterns and challenges in real user-agent collaboration workflows. Kimi K3 demonstrates consistent advantages across production-scenario-oriented workflows, indicating comprehensively enhanced agent knowledge work capabilities.

■ Research and Visualization

Below are examples of what K3 can accomplish in Kimi Work for financial advisory and research scenarios.

ASIC Industry Research: Covering 42 years of ASIC industry history, generated by K3 through over 120 rounds of recursive self-improvement. K3 transforms evidence into custom charts, animated diagrams, and interactive visual narratives. Throughout the process, it completed over 2,800 web searches and scrapes, over 1,100 terminal data pulls, and processed 87 quarterly reports and 99 original PDFs — totaling over 11,000 pages of material.

View the full interactive report: https://asic42cn.ok.kimi.link/

Controllable Nuclear Fusion Industry Research: A consulting-style industry study produced by K3, including timelines, tree diagrams, waterfall charts, Gantt charts, and presentation-quality slides.

GWTC-5 Gravitational Wave Analysis: An analysis of 391 gravitational wave events, using over 20 concurrent subagents, producing 7 scientific visualizations, 2 tables, and synthesizing literature from over 10 papers.

■ Widgets and Dashboards In Kimi Work, we are launching two new features: Widgets and Dashboard, making interactions with Kimi K3 more visual and persistent. Widgets can generate interactive components directly in conversation, connecting to local data or external plugins for continuous updates. Dashboards aggregate your most important widgets into a persistently saved personalized view, organized around a theme, project, or goal.

■ Video Editing Kimi K3 excels at motion design, animation, and video editing because its native multimodal architecture understands text, images, and video within a single model.

In one case, K3 produced a "3Blue1Brown"-style motion graphics explainer video about its own architecture, transforming technical concepts into animated diagrams and transitions — 4.5 minutes long:

In another case, Kimi K3 edited its own brand video from 56 raw clips, completing clip selection, motion-matched editing, frame-accurate beat matching, audio processing, and multiple revision rounds. A short video with this information density would typically take an experienced editor 1–2 workdays; a novice might need 3–5 days.

Architecture and Infrastructure Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA provides an efficient foundation for attention scaling, while AttnRes selectively retrieves representations across depth rather than uniformly accumulating them layer by layer. Together they form Kimi K3's architectural backbone, enabling the model to scale beyond trillion parameters.

Kimi K3 adopts Stable LatentMoE, actually activating 16 out of 896 experts. At this sparsity level, routing and optimization become critical challenges. Quantile Balancing directly assigns experts based on quantiles of routing scores, avoiding heuristic updates and sensitive balancing hyperparameters; Per-Head Muon extends Muon to optimize independently per attention head, making learning more adaptive in large-scale training. Sigmoid Tanh Unit (SiTU) and Gated MLA respectively enhance activation control and attention selectivity. These improvements collectively support stable, efficient training at the 2.8 trillion parameter scale.

Kimi K3 adopts quantization-aware training from the SFT stage, using MXFP4 weights and MXFP8 activations to accommodate broader hardware. To avoid throughput degradation from expert load imbalance in large-scale expert parallelism, we introduce fully balanced expert-parallel training using static shapes with no host synchronization on the critical path. Since inference efficiency also benefits from larger high-bandwidth communication domains, we recommend deploying Kimi K3 on supernode configurations of 64 or more accelerators. Finally, because KDA poses new challenges for traditional prefix caching, we have contributed the corresponding implementation to the vLLM community and will release it alongside the model. With KDA plus prefill cache, we can offer Kimi K3 at competitive token pricing even under large model scale and long context conditions.

More technical details will be published in a subsequent technical report.

Getting Started with Kimi K3 Using Kimi K3 Agents: Search for and download or upgrade to the latest version of the Kimi app in your mobile app store, supporting iOS, Android, and HarmonyOS; or visit kimi.com directly.

Working with Kimi K3: Download the latest version of the Kimi Work desktop client (version 3.1.0 and above), supporting Windows and Apple Silicon Mac computers.

Programming with Kimi K3: Kimi Code runs in your computer's terminal; select the K3 model via the /model command to start experiencing it.

Building applications with the Kimi API: Visit the Kimi API open platform, select kimi-k3. Pricing is 2 RMB per million tokens for input (cache hit) and 20 RMB (cache miss), and 100 RMB for output. Leveraging the Mooncake disaggregated inference architecture, official Kimi API programming scenarios achieve over 90% cache hit rate, making the effective input price just 1/4 of the standard input price. A top-up bonus of up to 30% is currently available.

Providing Kimi subscriptions for employees: Kimi Enterprise supports enterprise-grade data privacy protection and member management, with complete data isolation between personal and enterprise accounts. Click here to find the "Activate Enterprise Edition" button and subscribe to Kimi memberships for employees online.

Additionally, we are soon launching a hosted agent platform for enterprises: Kimi Hosted Agent, providing agent harnesses, isolated sandboxes, and long-task runtime environments. Join the waitlist to be among the first to participate in testing.

Limitations 1. Sensitivity to historical thinking content: Kimi K3 uses thinking history retention mode throughout post-training. If an agent framework fails to return all historical thinking content as required, or if switching to Kimi K3 from another model's ongoing session, contextual interference may occur, leading to unstable content generation quality. We recommend using agent frameworks that have passed compatibility verification such as Kimi Code, and avoiding switching to Kimi K3 mid-session. 2. Overly proactive: Kimi K3's training prioritizes long-horizon, high-difficulty tasks. Therefore, when encountering minor issues or ambiguous user intent during task execution, it may make unintended decisions on the user's behalf. If your application requires agents with stronger boundaries and less freelancing, please impose clearer behavioral constraints on Kimi K3 through system prompts or AGENTS.md. 3. Although Kimi K3 is overall a highly competitive model, there remains a certain gap in user experience compared to Claude Fable 5 and GPT-5.6 Sol.

Moonshot AI's Yutong Zhang Shares Her AI Entrepreneurship "Chain of Thought" for the First Time | Buming Startup Camp

Moonshot AI Releases and Open-Sources Kimi K2.6: Comprehensive Advances in Code, Long-Horizon Tasks, and Agent Swarm Capabilities | BlueRun Ventures Headlines

No Job Levels, No Departments, No OKRs: How 300 People at Moonshot AI Carry a 120 Billion Valuation | BlueRun Ventures Headlines