← Models

Kimi K2.6

Moonshot AI · Kimi

Moonshot AI's Kimi K2.6, a 1T-parameter MoE open-weight model for long-horizon coding and agentic workflows.

Part of Kimi family · Other versions: Kimi K2.5
Type
multimodal
Context
262K tokens
Max Output
33K tokens
Status
current
Input
$0.6/1M tok
Output
$2.5/1M tok
API Access
Yes
License
Modified MIT
chinese reasoning coding agentic open-weights multimodal long-context
Released April 2026 · Updated May 1, 2026

Overview

Freshness note: Model capabilities, limits, and pricing can change quickly. This profile is a point-in-time snapshot last verified on May 1, 2026.

Kimi K2.6 is Moonshot AI’s April 20, 2026 successor to K2.5. Moonshot positions it as a native-multimodal open-weight model built around long-horizon coding, autonomous tool use, and swarm-based task orchestration with up to 300 sub-agents and thousands of coordinated steps. It scales to 1 trillion total parameters with 32B active per token through a Mixture-of-Experts design and ships natively in INT4 quantization to keep self-hosted inference practical.

K2.6 keeps Moonshot’s pattern of releasing open weights alongside a managed API rather than gating frontier behavior behind a closed product.

Capabilities

Moonshot’s release materials highlight a specific capability profile:

  • Long-horizon agentic coding that sustains multi-step plans across long sessions, with reported strength on SWE-Bench Verified and Terminal-Bench style evaluations.
  • Native multimodality across text, images, and video in a single architecture, without separate vision modules.
  • Agent swarm orchestration scaling to hundreds of sub-agents and thousands of coordinated steps for distributed tasks.
  • Strong tool-calling behavior aligned with agent frameworks and managed agent harnesses.
  • Practical INT4 inference path that keeps the 1T-parameter model deployable on commodity GPU clusters.

Moonshot’s published benchmarks place K2.6 within striking distance of GPT-5.4 and Claude Opus 4.6 on agentic and coding tasks at a sharply lower price.

Technical Details

Public anchors at this snapshot:

  • 1T total parameters, 32B active per token, Mixture-of-Experts architecture.
  • 262K-token context window.
  • Native INT4 quantization for the open-weight release.
  • Native multimodal stack handling text, image, and video inputs.
  • Hosted API plus open weights on Hugging Face under a Modified MIT license.

The license functions as standard MIT for nearly all teams. Moonshot adds an attribution clause that requires displaying “Kimi K2” on the user interface only if the deployment exceeds 100 million monthly active users or generates more than $20 million in monthly revenue, with no usage restrictions below those thresholds.

Pricing & Access

Listed Moonshot API pricing (per 1M tokens):

  • Input: $0.60
  • Output: $2.50

Access options:

  • Moonshot API (kimi-k2.6)
  • Open-weight downloads on Hugging Face under Modified MIT
  • Third-party hosts such as Cloudflare Workers AI and OpenRouter
  • Local self-hosted inference via vLLM, SGLang, and similar runtimes

Best Use Cases

Choose Kimi K2.6 for:

  • Open-weight agentic coding workloads where self-hosting or full control is required.
  • Long-horizon coding agents that benefit from preserved planning state across thousands of steps.
  • Swarm-style orchestration where many sub-agents need to coordinate on subtasks.
  • Cost-sensitive production traffic that still needs reasoning-grade behavior.
  • Bilingual Chinese-English deployments with open-weight requirements.

For Western-distributed enterprise contracts, governance defaults, and broad cloud-platform availability, Anthropic, OpenAI, and Google remain easier operational baselines. K2.6 trades some of that for openness and cost.

Comparisons

  • Kimi K2.5 (Moonshot AI): Direct predecessor; K2.6 extends multimodality, swarm orchestration, and long-horizon coding behavior.
  • DeepSeek V4 (DeepSeek): Closest open-weight Chinese peer with similar 1M-context and aggressive pricing; V4 emphasizes long-context retrieval while K2.6 leans on agent orchestration.
  • Claude Opus 4.7 (Anthropic): Frontier closed-source alternative with stronger enterprise distribution; K2.6 competes on price and self-hostable open weights.