Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model featuring native vision capabilities and a 1-million-token context window. Positioned as the world's first open 3T-class model, Kimi K3 is available today through the Kimi API, Kimi.com, and the company's workspace apps. The full model weights are slated for release on July 27, 2026.
According to Moonshot AI, Kimi K3's overall performance still trails top-tier proprietary models like Claude Fable 5 and GPT 5.6 Sol. However, it demonstrated frontier-level capabilities in internal evaluations, particularly in long-horizon coding, complex knowledge work, and reasoning tasks, outperforming several other tested models including Opus 4.8 and GPT 5.5.
Under the Hood
Kimi K3 relies on structural changes to manage its massive size efficiently. It uses Kimi Delta Attention and Attention Residuals to improve how information flows through the model. Moonshot AI also scaled up Mixture of Experts (MoE) sparsity, effectively activating just 16 out of 896 experts per token using a framework they call Stable LatentMoE.
These architectural shifts yield roughly a 2.5x improvement in scaling efficiency compared to the previous generation, Kimi K2, allowing the model to convert compute power into useful output more effectively. At launch, Kimi K3 runs on a "max thinking effort" default, with low- and high-effort modes planned for future updates.
Proven in Deep Work
The technical report highlights Kimi K3’s capacity for sustained, autonomous work—what the developers call "long-horizon" performance. The model does not just generate snippets of text; it orchestrates terminal tools and navigates massive repositories with minimal human oversight.
In one test, Kimi K3 successfully built a compact GPU compiler (MiniTriton) from scratch, delivering performance on par with existing optimized stacks. In an early proof of concept for hardware engineering, the model spent a single 48-hour autonomous run designing, optimizing, and verifying a chip that successfully closed timing in simulation.
For scientific research, Kimi K3 completed an astrophysics computational workflow in two hours that normally takes an experienced researcher one to two weeks. The model autonomously reviewed more than 20 papers, cross-validated data, generated over 3,000 lines of Python code, and produced an interactive HTML dashboard.
What This Means for SMB Automation
Most small and mid-sized businesses do not need an AI to design a microchip or write GPU kernels. But the underlying mechanics required to perform those tasks—managing massive amounts of context, reasoning through multi-step problems, and executing long sequences of code autonomously—are exactly what businesses need to automate complex daily operations.
Here is how Kimi K3's specific capabilities map to practical business automation:
Massive context for messy operations. A 1-million-token context window means you can feed a model entire vendor manuals, years of customer service logs, or hundreds of pages of legal contracts in a single prompt. For an SMB, this reduces the need for expensive, complex database architectures just to get the AI to reference company policies when routing tickets or drafting vendor emails.
Long-horizon execution for back-office admin. Kimi K3’s ability to run autonomously for 48 hours proves its capacity for sustained "agentic" behavior. In an operations context, this translates to an AI that doesn't just read a static invoice. Instead, it can actively log into a portal, cross-reference the invoice against a purchase order in your ERP, spot a discrepancy, draft an email to the vendor, and update the accounts payable tracker—all without a human needing to prompt it at each individual step.
Vision-in-the-loop for legacy software. Kimi K3 integrates vision directly into its coding loop, allowing it to see live screenshots and refine its output instantly. Many SMBs rely on legacy software that lacks clean APIs. A model with native vision and strong coding capabilities can reliably drive browser-based automation, physically navigating clunky interfaces by looking at the screen rather than relying on brittle code scrapers.
Finally, Kimi K3’s sheer scale puts downward pressure on the broader AI market. As an open 3T-class model that aggressively competes with proprietary systems, it forces top-tier providers to keep API costs in check. For businesses building automated workflows, highly efficient, heavily capable models mean complex admin operations are getting cheaper and more reliable to deploy.