Kimi K3 and K2.5 Update: K3 Model, 1M Context, Agent Swarm, and What's New in 2026

Kimi had a busy 2026. Between January and July, Moonshot AI shipped three major model releases, crossed a $20 billion valuation, and briefly suspended new sign-ups because demand outpaced compute capacity. If you last checked in on Kimi a few months ago, a lot has changed. Here is everything worth knowing.

Kimi K3: The Big One

On July 16, 2026, Moonshot AI dropped the K3 model. The numbers are the headline: 2.8 trillion parameters, 1-million-token context window, native multimodal support, and a custom attention mechanism called KDA. Elon Musk replied "Impressive" to a benchmark post. Hugging Face co-founder Thomas Wolf asked if this was "another DeepSeek moment."

K3 is not a niche coding model or a text-only upgrade. It handles text, images, video, code, spreadsheets, and documents in a single model. On benchmarks, it hit #1 on Frontend Code Arena, placed third globally on Artificial Analysis's intelligence index, and scored higher than Claude Opus 4.8 and GPT-5.5 on real-world knowledge work tests.

The catch: demand crashed the servers. Three days after launch, Kimi paused new consumer subscriptions and redirected all compute to existing users. As of August 2026, the subscription freeze was still in effect while Moonshot scaled up infrastructure.

Kimi K2.5: Where 2026 Started

Released January 27, the moonshot ai kimi k2.5 model was what brought Kimi back into the global conversation after DeepSeek's dominance in early 2025. The kimi 2.5 ai model packed 1 trillion parameters in a Mixture of Experts architecture with 384 experts (8 active per token), a 256K context window, and the debut of MoonViT-3D, a 400-million-parameter vision encoder for native multimodal input. Developers can access fireworks ai kimi k2.5 through the Fireworks AI inference platform alongside OpenRouter and Together AI.

The standout feature was Agent Swarm, a beta capability that lets Kimi coordinate up to 100 specialized sub-agents working in parallel on complex tasks. It cut execution time by 4.5x and scored 50.2% on Humanity's Last Exam, beating GPT-5.2's 45.5% at 76% lower cost.

K2.5 also introduced four operating modes:

K2.5 Instant: Fast, no chain-of-thought responses
K2.5 Thinking: Deep reasoning with visible step-by-step logic
K2.5 Agent: Single-agent tool use for web search, code execution, file handling
K2.5 Agent Swarm: Multi-agent parallel task execution

It scored 96.1% on AIME 2025 and 83.1% on LiveCodeBench v6, well ahead of Claude Opus 4.5 at 64%. The model was open-sourced under a modified MIT license and is available on OpenRouter, Together AI, and NVIDIA NIM.

Kimi K2.6: The Coding Specialist

Between K2.5 and K3, Moonshot released Kimi K2.6, a coding-focused model that scored 60.4% on SWE-bench Verified, a benchmark that tests real-world software engineering tasks. It was also released via Fireworks AI, making it accessible through their inference platform.

K2.6 showed Moonshot's strategy of releasing specialized models alongside the flagship. The coding model handled debugging, refactoring, and full codebase analysis, complementing K2.5's broader capabilities.

Kimi K2: The Foundation

The kimi k2 ai model was the pre-2026 baseline and the foundation for everything that followed. The moonshot ai kimi k2 release established Moonshot's MoE architecture with 256 experts and MLA attention. The earlier kimi 2 ai generation had proven the concept, but K2 scaled it to 1 trillion parameters with a 128K context window.

What All These Updates Mean

If you chart Kimi's trajectory through 2026: K2.5 rebuilt its reputation after the DeepSeek shock, K2.6 proved it could compete on specialized coding benchmarks, and K3 put it in the top tier globally alongside GPT-5 and Claude Fable 5.

The trade-off is capacity. K3's compute demands are high enough that free-tier users sometimes hit wait times during peak hours. The subscription pause signals that Moonshot's infrastructure is playing catch-up with its model ambitions.

For users, the practical effect is that Kimi on Android and the web now runs a model competitive with anything from OpenAI or Anthropic, at the free tier. The K3 model is live in the Kimi app, and K2.5 remains available for lighter tasks through the API.

You can download Kimi on APKPure to get the latest version with K3 model access. Updates typically roll out within hours of new releases.