moonshotai/kimi-k2.6

Moonshot AI's frontier open model, built for long-horizon coding, agent swarms, and autonomous software engineering. 1 trillion parameters, 262k context window, vision and tool use.

175 runs
License
Playground API Examples README

Readme

Kimi K2.6

Kimi K2.6 is Moonshot AI’s open model for agentic coding and long-horizon task execution. It has 1 trillion total parameters and a 262,144 token context window, and it’s designed to run for hours at a time making tool calls, writing code, and coordinating with other agents.

What’s new in K2.6

Moonshot built K2.6 around three ideas:

Long-horizon coding. K2.6 can keep working on a single task for hours without losing its footing. In Moonshot’s showcase, it spent 13 hours overhauling an 8-year-old financial matching engine — 4,000+ lines of code, 12 optimization strategies, 1,000+ tool calls — and came out with a 185% throughput improvement. In another run it downloaded and deployed Qwen3.5-0.8B locally on a Mac, implemented inference in Zig, and tuned it from ~15 to ~193 tokens/sec across 14 iterations.

Coding-driven design. K2.6 turns short prompts into complete frontends with proper layouts, scroll-triggered animations, hero sections, and auth + database wiring. It’s fluent in Rust, Go, Python, TypeScript, and can use image and video generation tools to produce full landing pages.

Agent swarms. K2.6 can decompose a task into heterogeneous subtasks and run them concurrently across up to 300 sub-agents over 4,000 coordinated steps, producing documents, spreadsheets, slides, and websites in a single autonomous run.

Benchmarks

K2.6 posts frontier-level scores on agentic and coding benchmarks and consistently outperforms K2.5:

  • Terminal-Bench 2.0: 66.7
  • SWE-Bench Pro: 58.6
  • SWE-Bench Verified: 80.2
  • SWE-Bench Multilingual: 76.7
  • LiveCodeBench v6: 89.6
  • HLE-Full with tools: 54.0
  • BrowseComp: 83.2 (86.3 with agent swarm)
  • DeepSearchQA (f1): 92.5
  • GPQA-Diamond: 90.5
  • AIME 2026: 96.4
  • V* with python: 96.9

Using K2.6 on Replicate

This is a streaming chat model. Pass a prompt, optionally a system prompt, optionally images, and read the streamed output.

import replicate

for chunk in replicate.stream(
    "moonshotai/kimi-k2.6",
    input={
        "prompt": "Refactor this Python function to be idempotent and add tests.",
        "max_tokens": 4096,
        "reasoning_effort": "medium",
    },
):
    print(chunk, end="", flush=True)

Inputs

  • prompt — the user message.
  • system_prompt — sets the assistant’s behavior.
  • image_input — one or more images to include alongside the prompt. K2.6 is a vision-language model.
  • reasoning_effort — none (default, fast), low, medium, or high. Higher effort produces a longer thinking trace before the final answer.
  • max_tokens — up to 32,768 per response.
  • temperature, top_p, presence_penalty, frequency_penalty — standard sampling controls.

When to turn on reasoning

Reasoning is off by default. When you turn it on, K2.6 generates a hidden thinking trace before it answers, which eats into your max_tokens budget. For math, code, and multi-step problems, turn reasoning on and give it a generous max_tokens (4,000+). For short chat replies, leave it off.

Model created
Model updated