Moonshot/kimi-k3

Supported Functionality

Item	Specification
Input	Text, Image, Video (native vision)
Output	Text
Context	1,048,576 tokens (~1M)
Max Output	Up to 1,048,576 tokens (default 131,072)
Vision	✓ Supported
Function Calling	✓ Supported

Description

Kimi K3 is the flagship large language model released by Moonshot AI on July 16, 2026, succeeding the Kimi K2 family (K2, K2.5, K2.6, and K2.7 Code). It is a sparse Mixture-of-Experts model with 2.8 trillion total parameters, activating 16 of 896 experts per token, and is described by Moonshot as the world's first open 3T-class model, aimed at frontier intelligence across long-horizon coding, knowledge work, and reasoning. Full model weights are scheduled for release by July 27, 2026.

Its central breakthrough rests on two in-house architectural innovations: Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals (AttnRes), a drop-in replacement for residual connections. These improve how information flows across sequence length and model depth respectively, and combined with a Stable LatentMoE framework and refined training recipes, yield roughly a 2.5× improvement in scaling efficiency over Kimi K2. The model natively understands text, images, and video, and at launch runs an always-on reasoning mode set to max thinking effort by default.

Key Capabilities

• Long-horizon agentic coding: Sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools with minimal oversight — 88.3 on Terminal-Bench 2.1 and 42.0 on SWE Marathon (leading its tested peer set).
• Deep reasoning & knowledge: Always-on thinking supports complex chains of reasoning, scoring 93.5 on GPQA-Diamond and 43.5 on HLE-Full, built for multi-step tasks rather than isolated prompts.
• Native multimodal understanding: Unifies text, image, and video understanding in a single model — 81.6 on MMMU-Pro and 91.1 on OmniDocBench — enabling "vision in the loop" coding.
• Ultra-long context: A 1,048,576-token window processes enormous codebases and documents at once; reaches 90.4 on BrowseComp even at full window with no context management.
• Frontier engineering creation: Combines 3D reasoning, coding, and vision to turn concepts, images, and video into runnable interactive experiences (game dev, frontend, CAD).
• Agentic tool use: Native Tool Use / function calling and MCP support — 91.2 on BrowseComp — compatible with Kimi Code, Claude Code, Codex, and other harnesses.
• Efficient long-context inference: KDA delivers up to ~6.3× faster decoding in million-token contexts, and with a >90% cache-hit rate in coding workloads keeps serving costs competitive.

Technical Strengths

Feature	Benefit
Kimi Delta Attention (KDA)	Hybrid linear attention sharply speeds long-context decoding, keeping 1M-context inference affordable.
Attention Residuals (AttnRes)	Selective depth-wise retrieval delivers ~25% higher training efficiency at under 2% added cost, per Moonshot.
Stable LatentMoE sparsity	Activating 16 of 896 experts ties inference cost to active — not total — parameters, balancing scale and efficiency.
~2.5× scaling efficiency	Structural and training-recipe gains convert compute into intelligence more effectively at the 2.8T scale.
Quantization-aware training	MXFP4 weights with MXFP8 activations from the SFT stage onward ensure broad hardware compatibility.
Open weights + OpenAI-compatible API	Open-weight release and OpenAI SDK compatibility lower the integration and self-hosting barrier.

Capability Ratings

Dimension	Rating	Notes
Reasoning	Excellent	Backed by always-on thinking; 93.5 on GPQA-Diamond, though still behind top proprietary models on HLE.
Coding	Top-tier	Long-horizon coding is its headline strength, leading on SWE Marathon and Terminal-Bench.
Creative Writing	Strong	No official creative-writing benchmarks; Moonshot notes a UX gap versus top proprietary models.
Multimodal	Excellent	Native vision and video understanding with strong MMMU-Pro and OmniDocBench results.
Response Speed	Moderate	KDA accelerates long-context decoding, but always-on max thinking adds latency.
Context Window	Huge	1,048,576 tokens, among the largest currently available.

Use Cases

• Large-repository engineering: Multi-hour engineering sessions, cross-repo navigation, and terminal-tool orchestration under minimal supervision.
• Deep research & knowledge work: Combines large-scale retrieval with interactive visualizations to produce drill-down research reports and charts.
• Interactive software & game creation: Turns concepts and images into runnable 3D interactive experiences with vision-in-the-loop iteration.
• Multimodal document processing: Parses long documents, charts, and screenshots for structured understanding of filings and scientific literature.
• Agents & automation: Orchestrates multi-step automation and multi-agent collaboration via tool calling and MCP.
• Research reproduction: Bridges scientific literature and executable code to implement, validate, and analyze computational workflows.
• Enterprise private deployment: Kimi Enterprise provides data isolation and member management for team and organization use.

FAQ

What is Kimi K3 and who built it? Kimi K3 is the flagship large language model released by Moonshot AI on July 16, 2026. It succeeds the Kimi K2 family (K2, K2.5, K2.6, and K2.7 Code) and targets long-horizon coding, knowledge work, and reasoning.

How many parameters does Kimi K3 have? It has 2.8 trillion total parameters in a sparse Mixture-of-Experts design, activating 16 of 896 experts per token. Moonshot describes it as the world's first open 3T-class model.

What architecture does Kimi K3 use? It is built on two in-house innovations — Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals (AttnRes) — paired with a Stable LatentMoE framework, giving roughly a 2.5× improvement in scaling efficiency over Kimi K2.

What are Kimi K3's context window and max output? The context window is 1,048,576 tokens (~1M). max_completion_tokens defaults to 131,072 and can be set up to 1,048,576.

Does Kimi K3 support multimodal input? Yes. It natively accepts text, image, and video input with native visual understanding, and produces text output.

Is Kimi K3 open source, and when are the weights released? Moonshot positions it as an open model, with full model weights scheduled for release by July 27, 2026.

How is the Kimi K3 API priced? Flat pay-as-you-go with no tiering by context length: $0.30 per million tokens for cache-hit input, $3.00 for cache-miss input, and $15.00 for output.

How do I access Kimi K3? The API is OpenAI SDK-compatible with the model ID kimi-k3. It is also available through the Kimi app (iOS/Android/HarmonyOS), the Kimi Work desktop app, and Kimi Code.

How does Kimi K3's thinking mode work? K3 has thinking mode always on, configured via the reasoning_effort field. At launch only the max level is supported (default), with more levels to come.

How does Kimi K3 compare to top proprietary models? It reaches frontier-level performance across Moonshot's evaluation suite and consistently outperforms other tested models, while still trailing the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, overall.