Kimi

Overview

Kimi is Moonshot AI’s assistant, built around the K2.6 Mixture-of-Experts model — a 1-trillion-parameter network that activates about 32B per token — and engineered for long-horizon coding, agentic work, and document-heavy reasoning. Available at https://kimi.com, it pairs a 262K-token context window with an “agent swarm” that can spin up to 300 specialized sub-agents in parallel, and it ships open weights under a modified MIT license so teams can self-host via vLLM or SGLang. It is one of the few open-weight models our evaluation considers a genuine competitor to closed frontiers on engineering tasks.

What makes Kimi matter is that it closes the open-vs-closed gap on exactly the work developers care about. K2.6 leads SWE-Bench Pro at 58.6% and Terminal-Bench 2.0 at 66.7%, ahead of several closed rivals, and the 262K context means you can drop a whole textbook or codebase in and question it precisely. For bilingual (especially Chinese-English) research and document reasoning, it is excellent, and the agent swarm turns a single prompt into a coordinated multi-agent job.

The caveats our evaluation notes are honest trade-offs. Pure reasoning still trails the very top closed models on HLE (34.7% vs GPT-5.4’s 39.8%), and vision benchmarks lag leaders (MathVision 87.4% vs 92.0%). Most importantly, long autonomous runs consume heavy token volume, so API cost can spike on big tasks — the power is real but not free, and you must budget for it.

Key Features

• K2.6 MoE model (1T params, 32B active) — A large sparse model that reasons strongly on engineering while keeping per-token cost sane. This matters because it delivers frontier-ish code quality without a frontier-size bill on every call.
• 262K context window — You can ingest entire documents or codebases and ask targeted questions without chunking. For legal, academic, or multi-file work this is the headline feature.
• Agent Swarm orchestration — Kimi can launch up to 300 specialized sub-agents in parallel to divide and conquer a complex task, a differentiator for autonomous, long-horizon jobs.
• Design-to-code — It generates responsive React/Next.js from UI screenshots, turning a mockup into a working frontend faster than hand-coding.
• Open weights (MIT, modified) — Self-host via vLLM, SGLang, or TensorRT-LLM for privacy and cost control, rare among top engineering models.
• Native multimodal chat — Text, image, and document inputs are handled in one chat, supporting the bilingual and document-heavy workflows it is known for.

Pricing

In our evaluation the free tier is a fair on-ramp for chat and light document work. The $19–$39 tiers unlock the agent features that make Kimi distinctive, and the $99 tier is for heavy autonomous coding. The API is cheap per token but dangerous on long agentic runs — our advice is to cap task scope and watch usage, because the swarm’s power scales cost quickly. For privacy or volume, self-hosting the open weights fixes both price and data concerns.

How It Compares

vs. ChatGPT: ChatGPT is the broader, more polished generalist with a bigger ecosystem; Kimi wins on open-weight engineering benchmarks, 262K context, and the agent swarm. Our evaluation picks ChatGPT for general use, Kimi for long-context and agentic coding.

vs. Claude: Claude writes more naturally and is the steadier general assistant, but Kimi’s K2.6 leads on SWE-Bench-style engineering and offers self-hosting. For prose use Claude; for open-weight coding at scale, Kimi.

vs. Perplexity: Perplexity is a cited answer engine; Kimi is a reasoning and agentic assistant. For sourced lookup use Perplexity; for long-document reasoning and autonomous coding, Kimi.

Getting Started

• Open https://kimi.com and try the free tier — paste a long PDF or a codebase excerpt to feel the 262K context immediately.
• Use the ”@” file attachment to dump a whole document and ask per-section questions; lean on the long context instead of chunking manually.
• Try design-to-code by uploading a UI screenshot and asking for React/Next.js — it is a fast way to prototype a frontend.
• Experiment with the agent swarm on a bounded task first; watch credit and token use so a long run does not surprise you.
• Avoid the beginner mistake of unbounded autonomous runs — scope the task or costs spike; start small and scale up.
• Self-host the open weights via vLLM/SGLang if privacy or volume matters; it removes both the cap and the bill.
• Power-user tip: pair Kimi’s API with an editor like Cursor for in-editor agentic coding — the low input price makes always-on reasoning affordable if you cap scope.

Hands-on Verdict

Kimi (Moonshot) is our go-to for long-context document and bilingual work — it comfortably digests hundreds of pages and keeps citations, and the K2.6 model reasons well on math and code. In our evaluation it is a lighter, open-weight alternative to Perplexity for Chinese-language sourced queries and a stronger engineering play than many closed assistants, even if English prose trails Claude.

Who should use it: developers wanting open-weight, agentic coding; researchers and analysts with long documents; Chinese-English bilingual teams. Who should skip it: users needing best-in-class vision or pure reasoning (use ChatGPT), or a polished general chatbot. Practical tip: use the ”@” file attachment to dump a whole textbook and ask targeted questions per chapter — the 262K context is the feature, so lean on it instead of chunking manually, and cap agent scope to control token cost.