Kimi K2 in 2026: What to Use Now (K2.5, K2.6, K3)

Kimi K2 was a turning point for open-weight AI. When Moonshot AI released it in July 2025, it was the first serious open-source model purpose-built for agentic tasks: tool use, multi-step reasoning, code generation, not just conversation. With 1 trillion total parameters and 32 billion activated via a Mixture-of-Experts architecture, it competed directly with closed frontier models.

But a lot has changed since then. Groq dropped K2 entirely, successors have shipped vision, thinking, and MCP support, and the pricing landscape looks completely different. If you’re still thinking about Kimi K2 the way the July 2025 launch articles described it, you’re working with stale information. This article covers what K2 actually was, where the K-family stands in August 2026, and which model you should use today.

If you’re interested in building your own AI agent with Mastra or exploring other agentic frameworks, the K-family models are among the strongest open-weight options available.

What made Kimi K2 stand out?

Unlike traditional language models that excel at conversation, Kimi K2 was optimized for agentic tasks. It doesn’t just understand your request. It plans, executes, and delivers complete solutions using tools and multi-step reasoning. This made it the first open-weight model that could genuinely compete with closed models on tool-calling benchmarks.

Moonshot AI achieved this through a massive MoE architecture: 384 experts with 8 selected per token plus 1 shared expert, trained on 15.5 trillion tokens with zero training instability. The result was a model that scored 65.8% on SWE-bench Verified (single-attempt) and 70.6% on the Tau2 retail tool-use benchmark, matching closed frontier models at the time.

For developers looking at the best AI coding tools and agents, Kimi K2 was the model that proved open-weight agentic coding was viable.

Key technical specifications

Kimi K2 vs the field at launch (July 2025)

These comparisons reflect the competitive landscape when K2 launched. All three competitors listed here are now two or more generations behind — GPT-5.x, Claude Fable 5, and DeepSeek V3.2 are the current counterparts. For current-gen comparisons, see the K-family section below.

The K-family lineup in 2026: K2.5, K2.6, and K3

Every “Future Roadmap” item from the original K2 launch has shipped in a successor. Here’s the full timeline:

K2 proved open-weight agentic AI works, K2 Thinking added reasoning, K2.5 added vision, K2.6 pushed long-horizon coding, and K3 went frontier with 2.8 trillion parameters and a million-token context window.

If you’re comparing the K-family against other open-weight options, see our guide to the best open-source LLMs to replace Claude or GPT.

Kimi K2 pricing and availability today

Here’s the current state as of August 2026.

Original Kimi K2

The original K2 (0711 build) is still available on OpenRouter, but it’s now a legacy model served by a single provider:

• Price: $0.57/M input, $2.30/M output
• Context: 131K tokens, max output 100,352 tokens
• Throughput: ~34 tokens/second (at time of writing)
• Providers: 1 (single provider)

It works, but K2.5 is cheaper and better for most use cases.

What to actually use today

Self-hosting Kimi K2: the reality check

The original article mentioned “self-hosting available with vLLM/SGLang/KTransformers.” That’s technically true, but it needs context for the bitdoze audience.

Self-hosting engines (all confirmed in the official GitHub README):

What changed with successors: K2 Thinking added native INT4 quantization, and K3 introduced MXFP4 — both make smaller GPU clusters borderline feasible. But “borderline” still means dedicated GPU hardware, not a Hetzner VPS running your Docker stack.

License details: K2 through K2.6 are open-weight under a Modified MIT License. Free for most uses, but a separate commercial license is required if your product exceeds 100 million monthly active users or $20 million/month revenue. The license also has a synthetic-data carve-out (model outputs are not covered) — relevant if you’re fine-tuning. K3 uses a separate “Kimi K3 License.”

Kimi Vendor Verifier: Moonshot ships a tool to confirm a third-party endpoint is actually serving the right model. Worth using if you’re routing through less-known providers.

What shipped since launch: vision, thinking, MCP

The original K2 article had a “Future Roadmap” section. Everything in it landed:

Real-world testing with Kimi K2

I tested Kimi K2 by building a complete solar panel website using Astro in Zed editor. Here’s what happened:

Getting started: code examples

Updated code snippets for the models you should actually use today.

import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});

const completion = await openai.chat.completions.create({
  model: "moonshotai/kimi-k2.5",
  messages: [
    {
      role: "user",
      content: "Build a React component for a solar panel calculator",
    },
  ],
  tools: [
    {
      type: "function",
      function: {
        name: "create_file",
        description: "Create a new file with content",
      },
    },
  ],
});

console.log(completion.choices[0].message);

const completion = await openai.chat.completions.create({
  model: "moonshotai/kimi-k3",
  messages: [
    {
      role: "user",
      content: "Analyze this codebase and suggest architectural improvements",
    },
  ],
  // K3 supports reasoning_effort: "low" | "high" | "max" (default: "max")
});

curl -X POST "https://api.moonshot.ai/v1/chat/completions" \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {
        "role": "user",
        "content": "Help me build a complete web application"
      }
    ]
  }'

Performance and cost optimization

Technical innovation: MuonClip optimizer

Kimi K2’s stability comes from a training innovation that solved one of the biggest problems with large MoE models.

Key innovations:

• QK-Clip technique: Prevents attention logit explosions that cause training divergence in large MoE models
• Adaptive scaling: Dynamic adjustment based on attention patterns during training
• Stable training: Zero spikes during 15.5T-token training runs
• Token efficiency: Superior performance per training token compared to standard optimizers

MuonClip’s stability innovations carried forward into all K-family successors. The training stability it enabled is one reason Moonshot AI was able to iterate so quickly from K2 to K3 in just one year.

Final verdict: which Kimi model should you use?

Want to try Kimi K3 without the premium price? OpenCode Go gives you access to K3 and 17 other models for around $10/month. For a broader look at open-weight options, see our guide to the best open-source LLMs to replace Claude or GPT.

FAQ