Kimi K2.6

Name: Kimi K2.6
Author: Moonshot AI

Released

2026-04-20

Last refreshed

2026-10-08

Status

Researched 102d ago

Open sourceCommercial use: permittedMultimodalCodingRAGAgentsLong contextVisionClassificationJSON / Tool use

Moonshot AI releases · 6 in the last 12 months · this family litChangelog →

Specifications

Family: Kimi K2
Released: 2026-04-20
Context: 262k
Max output: 65,536
Parameters: 1T
Architecture: Mixture of Experts
Knowledge cutoff: 2025-04
Specialization: code
Openness: Open source
License: MITOSI-approvedCommercial use: permitted
Weights: Available·Moonshot Modified MIT License
Code: Available·Moonshot Modified MIT License

Created by

Moonshot AI

Lossless long-context AI innovation

Beijing, China

Founded 2023

Website

Pricing

Output / 1M

$3.40

Input / 1M

$0.800

Cheapest of 9 routes · Novita AI

Providers(9)

Cloudflare Workers AI NVIDIA NIM Moonshot AI Kimi Fireworks AI OpenRouter Microsoft Foundry Vercel AI Gateway Novita AI Together AI

View 9 provider routes

Links

Website HuggingFace

About

Kimi K2.6 is Moonshot AI's multimodal agentic coding model, released April 20 2026 under a Modified MIT license. Built on a 1-trillion-parameter MoE architecture (32B active, 384 experts with 8 selected per token plus 1 shared expert, 61 layers), it features a 262K context window and up to 65,536 output tokens. Supports native image and video inputs (screenshots, PDFs, spreadsheets). Designed for long-horizon coding with agent swarms of up to 300 sub-agents and 4,000 coordinated steps; Moonshot AI cites 200–300 sequential tool calls without task drift.

Top use-case fit: coding, agents, and build tasks

Coding

Q/$ D

Agents

Q/$ C

Provider price ladder

Compare all 9

Compare API pricing across 4 providers for input and output tokens, batch, and cached reads when available.

Provider	Input / 1M	Output / 1M	Cache	Route
Novita AI	$0.800	$3.40	-	Serverless
OpenRouter	$0.730	$3.49	-	Serverless
Cloudflare Workers AI	$0.950	$4.00	read $0.160	Serverless
Fireworks AI	$0.950	$4.00	-	Serverless

Available via routers & gateways(7)

LiteLLM

Gateway

Open-source Python SDK and proxy server that unifies 100+ LLM APIs behind a single OpenAI-compatible interface, with load balancing, cost tracking, and configurable failover.

Free OSSMicrosoft Foundry

OpenRouter

Hybrid

Unified hybrid gateway to 400+ models from 60+ providers via a single OpenAI-compatible API, with optional auto-routing that selects the best model per prompt.

PassthroughTogether AIFireworks AI

Portkey

Gateway

Production AI gateway routing to 1,600+ LLMs with failover, load balancing, semantic caching, and guardrails; Apache 2.0 core is fully self-hostable with the complete feature set.

SubscriptionMicrosoft Foundry

Azure AI Foundry Model Router

Router

Microsoft Azure AI Foundry's native model router that uses a trained ML model to route each prompt in real time to the optimal Azure-hosted model, with Balanced/Cost/Quality mode selection and automatic failover.

PassthroughMicrosoft Foundry

Helicone

Gateway

Observability-first AI gateway with routing, caching, rate limiting, and request tracing; Apache 2.0 open-source core with a managed hosted tier for logging and analytics.

SubscriptionMicrosoft Foundry

Kong AI Gateway

Gateway

Multi-LLM AI gateway built on Kong Gateway 3.x, adding semantic routing, load balancing, guardrails, and MCP traffic analytics as plugins over Kong's existing API management platform.

SubscriptionMicrosoft Foundry

Capabilities

VisionMultimodalReasoningJSON / Tool useStructured OutputsPrompt Caching

Benchmark peer barsfor Coding

SWE-bench ProRank 16 of 49

80.3

79.2

69.2

67.7

58.6

SWE-bench VerifiedRank 16 of 92

Claude Fable 5

96.0

Claude Opus 5

96.0

Claude Mythos Preview

93.9

Claude Opus 4.8

88.6

Kimi K2.6current

80.2

HumanEvalRank 12 of 99

Claude Sonnet 4.6

98.0

96.7

Claude Opus 4.6

95.0

Grok-3

94.5

Kimi K2.6current

92.0

Benchmark scores(17)

Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.

Benchmark	Score	Version	Evaluation	Source
SWE-bench Verified	80.2	SWE-bench VerifiedObserved 2026-04-24	—	Source
Chatbot Arena	1462.0	—Observed 2026-05-17	—	Source
Google-Proof Q&A	90.5	GPQA Diamond (accuracy)Observed 2026-06-07	—	Source
MMLU PRO	84.6	From Kimi K2 (accuracy)Observed 2026-06-07	—	Source
HumanEval	92.0	—Observed 2026-04-29	—	Source
SWE-bench Pro	58.6	SWE-bench Pro (resolved%)Observed 2026-06-07	—	Source
Instruction-Following Evaluation	89.8	—Observed 2026-04-29	—	Source
MMMU Pro	80.1	LLM-Stats aggregatorObserved 2026-06-07	—	Source
LiveCodeBench	89.6	v6Observed 2026-04-20	—	Source
Terminal-Bench 2.0	66.7	Terminal-Bench 2.0 (accuracy%)Observed 2026-06-07	—	Source
SWE-bench Multilingual	76.7	—Observed 2026-04-20	—	Source
AIME 2026	96.4	AIME 2026 (accuracy)Observed 2026-06-07	—	Source
BrowseComp	83.2	BrowseComp (accuracy%)Observed 2026-06-07	—	Source
Humanity's Last Exam	34.7	HLE-Full without tools (accuracy)Observed 2026-06-07	—	Source
MCP-Atlas	55.9	MCP-Atlas (accuracy%)Observed 2026-06-07	—	Source
CursorBench	47.6	CursorBench 3.1Observed 2026-06-30	Configuration: Kimi 2.6 (single reported configuration) Harness: CursorBench 3.1 Evaluator: Cursor Confidence: confirmed Notes: Cursor published one CursorBench 3.1 configuration for this model; no cross-effort selection was needed.	Source
GeneBench-Pro	4.4	xhighObserved 2026-06-30	—	Source