Kimi K3 Complete Guide: Moonshot’s 2.8T Open-Weight Frontier Model (2026)

Moonshot AI just dropped something that reshapes the entire AI landscape. Kimi K3, released July 16, 2026, is a 2.8 trillion parameter open-weight model that scores 88.3% on Terminal-Bench 2.1. Only GPT-5.6 Sol beats it, and barely, at 88.8%.

This is not an incremental upgrade from K2.7. This is a generational leap that puts an open-weight model firmly in the top 3 globally.

What Makes Kimi K3 Different

K3 introduces Stable LatentMoE, a new mixture-of-experts architecture with 896 experts that activates only 16 at inference time. This gives you the reasoning capacity of a 2.8T model with the latency characteristics of something much smaller.

The key specs:

• Parameters: 2.8 trillion total, with 16 of 896 experts active per forward pass
• Architecture: Stable LatentMoE (novel approach to expert routing)
• Context window: 1 million tokens
• Vision: Native multimodal vision support
• License: Open-weight (available for commercial use)
• API pricing: $3 input / $15 output per million tokens
• Cache pricing: $0.30 per million tokens with 90%+ hit rate

The 1M context window with native vision means this model can process entire codebases while also understanding screenshots, diagrams, and visual documentation.

Benchmark Performance

Here is where K3 gets genuinely surprising.

Terminal-Bench 2.1: 88.3% (Rank #2 Global)

Terminal-Bench measures real-world terminal task completion. K3 scores 88.3%, placing it second globally behind only Sol at 88.8%. For context, Claude Opus 4.8 scores significantly lower on this benchmark. This is the first time an open-weight model has claimed a top-2 position on Terminal-Bench.

DeepSWE: 67.5%

On the DeepSWE benchmark for real software engineering tasks, K3 hits 67.5%. This places it above Opus 4.8 and well above Sonnet 5 at 63.2%.

Frontend Code Arena: #1

K3 takes the top spot in Frontend Code Arena, beating every other model at generating frontend code. If you build web interfaces, this matters a lot.

Intelligence Index: 57.1 (Rank #3)

On the broader Intelligence Index that measures general reasoning, K3 ranks #3 globally with 57.1. It beats GPT-5.5 and sits just under Fable 5 and Sol.

Architecture Deep Dive: Stable LatentMoE

Traditional Mixture-of-Experts models suffer from routing instability. Experts can become underutilized or overloaded, leading to degraded performance. Moonshot’s Stable LatentMoE addresses this with a new routing mechanism.

With 896 total experts and only 16 active per token, K3 achieves extreme sparsity. Each forward pass uses roughly 1.8% of total parameters. This keeps inference costs manageable despite the massive parameter count.

The “Stable” part refers to the training methodology. Moonshot claims their routing algorithm prevents the expert collapse problem that has plagued other MoE models at scale. The result is more consistent quality across different types of tasks.

The “Latent” aspect involves how expert representations are stored and accessed. Rather than having fully separate expert networks, K3 uses a latent space approach where experts share a compressed representation space. This reduces memory requirements for deployment.

Pricing and Access

K3 is available through OpenRouter as moonshotai/kimi-k3 and directly through Moonshot’s API.

	Input	Output	Cache
Kimi K3	$3/M	$15/M	$0.30/M
Claude Opus 4.8	$5/M	$25/M	varies
GPT-5.6 Sol	$5/M	$30/M	varies
Claude Sonnet 5	$2/M	$10/M	varies

The real pricing story is the cache. K3 achieves 90%+ cache hit rates in typical coding workflows. That means most of your input tokens cost $0.30 per million instead of $3. For long-context coding sessions where you repeatedly reference the same codebase, effective input costs drop by 90%.

For a full pricing breakdown, see our AI API pricing comparison.

Best Use Cases for Kimi K3

Agentic Coding

K3 excels at multi-step coding tasks that require terminal interaction. Its 88.3% Terminal-Bench score means it handles complex build systems, debugging sessions, and deployment workflows with near-perfect reliability.

You can use K3 with Aider or in Claude Code workflows through OpenRouter. See our setup guide for configuration details.

Frontend Development

Being #1 on Frontend Code Arena is not a fluke. K3 produces clean, modern frontend code with proper component architecture. React, Vue, Svelte, whatever your framework, K3 handles it better than any other model right now.

Large Codebase Analysis

The 1M context window means you can load substantial portions of a codebase for analysis, refactoring, or documentation. Combined with native vision, you can include architecture diagrams alongside code.

Open-Weight Deployment

Because K3 is open-weight, organizations can self-host. This matters for companies with strict data residency requirements or those who want to fine-tune on proprietary codebases.

How K3 Compares to the Competition

K3 sits in a very specific tier. It beats Opus 4.8 and GPT-5.5 across most benchmarks. It loses narrowly to Sol on Terminal-Bench (88.3% vs 88.8%) and to Fable 5 on certain reasoning tasks.

Against other open-weight models, the gap is enormous. DeepSeek V4 Pro scores 62.1% on comparable benchmarks. Tencent Hy3 leads on SWE-Verified at 74.4% but does not match K3 on Terminal-Bench or the Intelligence Index.

For the full comparison with Chinese models, see our best Chinese AI models guide and our best open-source coding models roundup.

Limitations

K3 is not perfect. A few things to watch:

• Cost at scale: At $15/M output tokens, extended generation gets expensive. The cache helps with input but output pricing is premium-tier.
• Availability: While open-weight, deploying 2.8T parameters requires significant infrastructure. Most users will access it through APIs.
• Recency: The training cutoff means it will not know about very recent libraries or frameworks released after its training window.
• Not the best at everything: Sol still edges it on Terminal-Bench. Hy3 has better SWE-Verified scores. K3 is a generalist that excels broadly rather than dominating every niche.

Getting Started

The fastest path to K3 is through OpenRouter:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "messages": [{"role": "user", "content": "Hello K3"}]
  }'

For detailed API setup, caching optimization, and code examples, read our Kimi K3 API tutorial.

Who Should Use Kimi K3

• Teams wanting frontier performance without vendor lock-in: Open-weight means you can always self-host if pricing changes.
• Frontend developers: Best-in-class frontend code generation.
• Agentic coding workflows: Near-Sol-level terminal task completion.
• Budget-conscious teams: Cache hits at $0.30/M make repeated interactions very cheap.
• Organizations building on Chinese AI: K3 represents the current peak of the Chinese AI ecosystem for coding.

FAQ

Is Kimi K3 better than Claude Opus 4.8?

On benchmarks, yes. K3 scores higher on Terminal-Bench 2.1 (88.3% vs Opus 4.8’s lower score), DeepSWE (67.5%), and the Intelligence Index (57.1 vs Opus). The gap is meaningful. However, Opus 4.8 still has strengths in certain instruction-following and safety-critical tasks. See our full K3 vs Opus 4.8 comparison.

Is Kimi K3 open source?

K3 is open-weight, meaning the model weights are available for download and commercial use. The training code and data are not open-source. This is the same approach used by Llama and other “open” models.

How does the cache pricing work?

K3 uses prefix caching. When you send a request that shares a prefix with a previous request (common in coding where you repeatedly send the same system prompt and codebase context), those shared tokens are served from cache at $0.30/M instead of $3/M. Moonshot reports 90%+ hit rates in typical coding workflows.

Can I run Kimi K3 locally?

Technically yes, since it is open-weight. Practically, 2.8T parameters require multiple high-end GPUs. Most individual developers will use the API. Enterprise teams with dedicated GPU clusters can self-host.

How does K3 compare to Sol?

Sol beats K3 on Terminal-Bench by 0.5 percentage points (88.8% vs 88.3%). Sol also costs more ($5/$30 vs $3/$15). For most coding tasks, the performance difference is negligible, but Sol maintains a slight edge at the very top. See our detailed K3 vs Sol comparison.