Is the Kimi Hype Over? Why Moonshot's Kimi Is Still Useful

For a while in 2025, Kimi was the model everyone talked about. Moonshot AI's Kimi K2 became the most downloaded model on Hugging Face the day after its release, and later versions were compared with the best closed models at a fraction of the price.

In 2026 the mood shifted. Headlines moved from benchmarks to controversy, and newer models stole attention. So is Kimi still worth your time? Short answer: yes, for specific jobs. Here's the honest picture.

A quick timeline

When	Release	Why it mattered
Jul 2025	K2	1 trillion parameters (32 billion active), open weights, modified MIT licence. Viral on Hugging Face.
Nov 2025	K2 Thinking	Reasoning version that could chain 200 to 300 tool calls. Reported training cost around $4.6 million.
Jan 2026	K2.5	Added vision, "Agent Swarm" with up to 100 sub-agents, and the Kimi Code CLI.
16 Jul 2026	K3	2.8 trillion parameters, open weights, the largest open model reported. Called a "DeepSeek moment" by some outlets.

Why the hype cooled

Controversy. In February 2026, Anthropic accused Moonshot, DeepSeek and MiniMax of a "distillation attack" using fraudulent accounts. In September, Anthropic accused Moonshot specifically of routing requests to Claude through thousands of proxy accounts and training on the responses. In April, a US House committee questioned companies using Kimi models. These are allegations; they don't change what the downloadable model does, but they make some companies cautious.

Competition. Since July, the leaderboard has moved. Epoch AI's capability index ranked K3 8th of 223 models at release; by October it sat 19th of 253. Still near the top of open models, no longer near the top overall.

Price. K3's API costs around $3 per million input tokens and $15 per million output tokens. That's mid-range, not bargain-basement. If cheap is all you need, models like GPT-6 Luna, Claude Haiku 5.5 or DeepSeek's Flash cost a fraction of that. (See our AI model price comparison.)

Why Kimi is still useful

1. Open weights you can own

This is the big one. You can download K3 and run it on your own hardware or with a provider you choose. That means:

• Your data stays where you decide. Useful for regulated industries, or anyone uneasy about sending data abroad.
• You can fine-tune it on your own domain.
• No vendor can switch it off or change it under you. Pin a version and it behaves the same forever.

Closed models, however good, can't offer that.

2. It's strong where it counts

K3 scored 93% on GPQA Diamond (graduate-level science questions) and 97% on a mock AIME maths exam in Epoch AI's tracking. At release it outperformed Claude Opus 4.8 and GPT-5.5 on several benchmarks. For an open model you control, that's very capable.

3. Built for agents

Kimi's line has focused on long agent runs: K2 Thinking chained hundreds of tool calls, and K3's Agent Swarm runs up to 300 sub-agents. If you're building agents that work through many steps, it's a serious option to test.

4. Others are building on it

Several products have reportedly built on Kimi models, including Cursor's Composer 2 (reported to be based on K2.5). When coding-tool makers choose a base model, it's usually because it performs well and they can customise it.

The licence catch

K3's licence isn't fully open in the way K2's modified MIT licence was. Inference providers earning more than $20 million a year must sign a commercial agreement with Moonshot, with revenue sharing of up to 30%. For most companies running it for their own use, that won't apply, but read the licence before building a business on top of it.

So, should you use it?

Your situation	Kimi K3?
You need data to stay on your own servers	Strong candidate
You want to fine-tune an open model	Strong candidate
You're building multi-step agents	Worth testing
You just want the cheapest API	Look elsewhere
Strict rules about China-linked vendors	Check with compliance; self-hosting may still be allowed

The hype is over, and that's fine. Hype is a poor way to choose tools anyway. What's left is a very capable open-weights model with a clear set of jobs it's good at.

In one line: Kimi is no longer the headline, but K3's open weights, strong reasoning and agent focus still make it a smart pick when you need control over where your model runs.