Perplexity just flipped the switch on Kimi K2.6, letting Pro and Max subscribers tap into Moonshot’s new state-of-the-art open-weight model right from the model selector, no setup, keys, or infra drama required.

Kimi K2.6 is Moonshot AI’s latest flagship open-weight model, built as a mixture-of-experts system with 1T total parameters and 32B active at inference, continuing the same architecture line as Kimi K2 and K2.5. It’s trained to excel at long-horizon coding, agent-style workflows, and complex reasoning, which makes it a strong fit for the kind of deep research, code refactoring, and multi-step planning people already lean on Perplexity for.

On paper, this thing is no lightweight: independent evaluations put Kimi K2.6 as the top open-weights model globally, ranking fourth overall on Artificial Analysis’ Intelligence Index and coming within just a few points of frontier closed models from the big labs. Benchmarks highlight a big jump in “agentic” performance – think preparing long reports, orchestrating tools, or managing multi-step tasks – with its Elo score on a general agentic eval climbing from 1309 in K2.5 to 1520 in K2.6. In practice, that translates to fewer derails mid-task and more “it just kept going and got it done” moments when you throw long, messy workflows at it.

Crucially for an AI search and research app, Kimi K2.6 also focuses on staying grounded: Artificial Analysis reports a much lower hallucination rate than its predecessor, putting it in the same ballpark as premium proprietary models like Claude Opus on their knowledge and accuracy tests. For Perplexity users, that means you’re getting an open-weight model that’s not just powerful, but also better at saying “I don’t know” instead of confidently making things up when sources are thin.

From the open-source side of the story, K2.6 is a big deal because Moonshot released it as open weights from day one, with the model hosted on platforms like Hugging Face for anyone to integrate, fine-tune, or self-host. It posts strong coding numbers too, with external writeups citing SWE-Bench scores that put it in “good enough for serious engineering workflows” territory, not just toy coding demos. That combination – open weights plus near-frontier performance – is exactly why people are calling it a watershed moment for open models.

Perplexity plugging Kimi K2.6 directly into Pro and Max is where it gets interesting for everyday users rather than just model nerds. Instead of having to wrangle APIs or manage your own stack, you can now pick Kimi K2.6 as an option alongside other frontier models, and let Perplexity handle routing and orchestration under the hood. For heavy Pro users, Kimi’s strength in long-horizon coding and agent-like tasks should pair nicely with Perplexity’s existing “Thinking” style and computer features, especially on big research projects, documentation reworks, or complex coding prompts.

Zooming out, this move also says a lot about where the ecosystem is heading: a model orchestration product like Perplexity choosing to feature an open-weight model that’s closing the gap with closed systems makes open AI feel a lot less “second tier.” If Kimi K2.6 performs as well in real user workloads as it does in benchmarks, expect more people to start asking why their daily tools can’t ship competitive open-weight defaults too – and more stack providers to follow this playbook.