Moonshot Kimi LLM Series

by Moonshot AI

Kimi is Moonshot AI’s flagship model family, spanning K1 (Nov 2023) through K3 (July 2026). K3 is frontier-class with 2.8T parameters, 1M context, native multimodal, open weights, and Agent Swarm capabilities. Emphasizes agentic autonomy, long-context reasoning, and cost-efficiency vs Claude/GPT.

Overview

Kimi represents China’s major push into frontier AI. The model family progresses from baseline reasoning (K1) to a comprehensive agentic platform (K3), emphasizing open-source releases, long-horizon task execution, and competitive positioning against Claude and GPT on cost-per-task and agentic performance.

Company Context: Moonshot AI, founded 2023 by Tsinghua grad Yang Zhilin; backed by Alibaba and Chinese internet giants; achieved unicorn status ($1B valuation).

Complete Model Lineage

Model	Release	Parameters	Key Feature	Status
K1	Nov 16, 2023	Undisclosed	Baseline reasoning	Archived
K1.5	Jan 20, 2025	Undisclosed	o1-level reasoning	Limited
K2	July 11, 2025	1T / 32B active	384-expert MoE, open-weight	Active
K2 Linear	Oct 2025	48B / 3B active	Lightweight KDA	Active
K2.5	Jan 2026	1T / 32B active	Multimodal + 100 agents	Active
K2.6	April 2026	1T / 32B active	300 agents, 4K steps, #1 open-weight	Active
K2.7 Code	June 2026	1T / 32B active	+21.8% coding, -30% tokens	Active
K3	July 16, 2026	2.8T / ~50B active	Frontier: 1M context, KDA, always-on reasoning, open weights	Active

K3 (Current Flagship) — Detailed Specifications

Architecture

Parameters & Scaling:

• Total: 2.8 trillion parameters (896 experts)
• Active per token: ~50 billion (16 active experts via learned routing)
• Attention: Hybrid KDA (Kimi Delta Attention) — 3:1 ratio of linear to full attention
• Efficiency gain: 2.5× better compute-to-intelligence vs K2
• Quantization: MXFP4 weights, MXFP8 activations (quantization-aware training)
• Attention Residual (AttnRes): Improved information flow across layers

Capabilities

Context & Speed:

• Context window: 1 million tokens (~750,000 words)
• Max output tokens: ~64,000
• Decoding speed: 6.3× faster than K2 for long-context queries
• Reasoning mode: Always-on by default (can disable for speed)

Multimodal:

• Native vision understanding via MoonViT-3D encoder
• Image + text reasoning at 1M context (unique capability)
• Chart, diagram, and document analysis

Agent Swarm:

• 300+ parallel sub-agents
• 4,000+ coordinated reasoning steps
• 4.5× faster execution vs single-agent approach
• Multi-path exploration and consensus-building

Performance & Benchmarks

Agentic/Autonomous Tasks:

• Arena.ai Frontend Code: Rank #1 (beats Claude/GPT on web dev)
• Agentic Benchmarks: ~89.5% (5th most capable frontier model)
• Real-world task wins: 5 of 6 vs Claude Sonnet 5

Coding Performance:

• SWE-Bench Pro (via K2.6): 58.6%
• Terminal-Bench 2.0: 66.7%
• Internal coding benchmarks: +21.8% vs K2.6 (via K2.7 Code)

General Intelligence:

• Artificial Analysis Index: Rank #3 overall (behind Fable 5 ~60, GPT-5.6 Sol ~59)
• Math/Reasoning: Frontier-level performance
• Knowledge: Current through April 2026

Pricing & Availability

API Pricing (July 2026):

• Input: $3permilliontokens($0.30 with cache-hit)
• Output: $15 per million tokens
• Cost vs alternatives: 50-65% cheaper per completed task than Claude Sonnet 5

Subscription Status:

• Paused new subscriptions July 17, 2026 (overwhelming demand)
• Existing subscribers unaffected
• Capacity reopening batched as infrastructure scales
• Hosted API: platform.kimi.ai
• Web Chat: kimi.com

Open Weights:

• Released July 27, 2026 (Modified MIT license)
• Available on Hugging Face
• vLLM support announced for self-hosting
• Enables local deployment without cloud dependency

K2.6 (Agentic Production Model) — Specifications

Key Features

Agent Swarm 2.0:

• 300 parallel sub-agents (vs K2.5’s 100)
• 4,000 coordinated reasoning steps
• 4.5× faster execution on large-scale search tasks
• 3-4.5× reduction in critical path steps

Performance:

• Artificial Analysis (April 2026): Top ranked open-weight model
• Benchmarks: 58.6% SWE-Bench Pro, 66.7% Terminal-Bench 2.0
• Multimodal: Native vision via MoonViT-3D
• Context: 256K tokens

Pricing:

• Input: $0.60 per million tokens
• Output: $2.50 per million tokens
• Cost advantage: 7-10× cheaper than K3 for standard tasks

Use Cases

• Production agentic workflows (parallel research, coordinated reasoning)
• Cost-optimized deployments requiring open weights
• Coding assistance and software engineering
• Multi-agent systems

K2.5 — Multimodal + Agent Introduction

Release: January 2026

Features:

• First Kimi with native multimodal (vision + text)
• Agent Swarm 1.0: up to 100 parallel sub-agents
• 256K context window
• Open-source (Modified MIT)
• Instant & thinking modes

Positioning: Entry point to Kimi’s agentic capabilities; more capable than K2 but less scaled than K2.6.

K2 (Open-Weight Baseline) — Specifications

Release: July 2025

Architecture:

• 384-expert MoE (1T total, 32B active)
• Pre-trained on 15.5 trillion tokens
• Text-only (no vision)
• Open weights (Modified MIT)

Purpose: Credibility reset — released open-source to restore developer trust after consumer chatbot popularity declined.

Availability: Open-weight; self-hostable; lower pricing tier.

K2 Linear — Edge Deployment Variant

Release: October 2025

Specifications:

• 48B total, 3B active parameters
• KDA (Kimi Delta Attention) for efficiency
• Reduced memory footprint
• Fast inference

Use Case: Memory-constrained edge devices, cost-optimized inference at scale.

Competitive Positioning

vs. Claude Sonnet 5

Metric	Winner	K3	Sonnet 5
Intelligence	Claude	57/100	53/100
Agentic Performance	Kimi	89.5	81.9
Generation Speed	Claude	32.1 tok/s	76.1 tok/s
Context Window	Kimi	1M	256K
Price/Task	Kimi	-50-65%	baseline
Open Weights	Kimi	Yes	No
Real-world Wins	Kimi	5 of 6	1 of 6

Summary: Kimi K3 wins on agentic autonomy, context, cost-per-task, and openness. Claude wins on raw speed.

vs. GPT-5.6 Sol

Metric	Winner
Frontier reasoning	GPT Sol
Agentic coding	Kimi K3
Cost-per-task	Kimi K3
Open deployment	Kimi K3
Ecosystem maturity	GPT
Speed	GPT

vs. DeepSeek V4 Pro

Metric	Winner
Price-per-token	DeepSeek
Agent Swarm	Kimi
Multimodal	Kimi
Reasoning depth	Comparable
1M context	Comparable

Unique Differentiators

1. Agent Swarm (Moonshot-Exclusive)

• Parallel sub-agent orchestration (300 agents, 4K steps)
• Enables multi-path reasoning and exploration
• 4.5× faster execution for large-scale searches
• Unique capability vs Claude and GPT

2. Open Weights Strategy

• K3, K2.6, K2.5, K2 all open or promised open
• Differentiates vs Claude/GPT (closed)
• Addresses Chinese chip supply constraints (enables broader adoption without vendor lock-in)
• Enables private/local deployment

3. Long Context + Multimodal Native

• 1M token context with vision reasoning (K3)
• Full-codebase analysis, long-document reasoning
• Vision understanding integrated, not bolted-on

4. Cost-Per-Task Advantage

• 50-65% cheaper than Claude for complete task execution
• K2.6 at $0.60/$2.50 undercuts most competitors
• DeepSeek cheaper per-token, but Kimi wins per-task due to efficiency

Chinese Market Context

Market Trajectory

• Aug 2024: Kimi chatbot ranked #3 in China (monthly active users)
• June 2026: Dropped to #7 after DeepSeek R1 disruption (Jan 2026)
• Strategic pivot: Yang suspended chatbot marketing; pivoted to enterprise + developer focus
• K3 release: Aims to reclaim “Chinese AI innovation leadership” narrative

Strategic Differentiation

1. Hardware Constraints: US export controls limit chip access → open models enable adoption without vendor dependency
2. Agent-First: Targets enterprise workflow automation and developer productivity
3. Cost Leadership: Competitive on effective cost-per-task despite DeepSeek’s per-token pricing
4. Chinese-Language Optimization: Built-in multilingual capability for Chinese market

Competitive Dynamics

• DeepSeek disrupted pricing expectations (Jan 2026) with ultra-low per-token costs
• Moonshot responds with Agent Swarm differentiation + performance edge on agentic tasks
• Both companies emphasize open weights (vs. US closed-model strategies)

When to Use Each Kimi Model

Use K3 for:

• Frontier-class reasoning and autonomous task execution
• Long-context document/codebase analysis (1M tokens)
• Agentic workflows requiring parallel sub-agent coordination
• Open-weight deployment without cloud dependency
• Cost-per-task optimization on complex problems

Use K2.6 for:

• Production agentic systems with moderate budget
• Parallel research and multi-step reasoning
• Open-source applications
• Cost-optimized deployments

Use K2.5 for:

• Multimodal reasoning with agent support
• Entry to Kimi’s agentic capabilities
• Vision + text tasks

Use K2 for:

• Open-source, cost-sensitive deployments
• Text-only reasoning

Use K2 Linear for:

• Edge/embedded deployment
• Memory-constrained environments

See Also

• Kimi K2 Thinking
• Kimi + OpenClaw Agent
• Claude Fable 5
• GPT-5.6 Sol
• Moonshot AI
• AI Agentic Systems