What Is Kimi K3? Moonshot AI's 2.8T Open-Weight Model Explained

On July 16, 2026, Moonshot AI released a model with 2.8 trillion parameters — the largest open-weight model ever shipped by that count. Its benchmark score puts it on par with Claude Opus 4.8. Its output token price is $15 per million. Its open weights arrive eleven days later, on July 27.

That's Kimi K3 in four numbers. Here's what they mean for developers and teams deciding whether to use it.

What this article covers: K3's architecture and why 2.8T parameters doesn't mean 2.8T compute, how it benchmarks against frontier models, how to call the API today, what it costs, when it's worth it, and when a cheaper model like GLM 5.2 is the better fit.

What Makes Kimi K3 Different

Two things set Kimi K3 apart from previous open-weight models.

First, scale. 2.8 trillion parameters makes K3 the largest open-weight model released as of July 2026 — larger than Llama 405B by a factor of nearly seven, and approaching the estimated parameter count of proprietary frontier models. That scale corresponds to K3 scoring 57 on the Artificial Analysis Intelligence Index, placing it in the same tier as Claude Opus 4.8 and GPT-5.5.

Second, efficiency. K3 is a Mixture-of-Experts (MoE) model with 896 experts per layer. Only 16 of those 896 are activated for any given token — about 1.8% of the pool. The 2.8T number describes the full parameter count; the active compute at inference time is a fraction of that. This is the same architectural trick used by Mixtral and DeepSeek — enormous parameter space for knowledge and capability, efficient forward pass in practice.

The combination of large parameter space and sparse activation is why K3 can achieve frontier-tier benchmark scores while remaining a candidate for open-weight release. It's not the same as running a dense 2.8T model.

Key Specs at a Glance

Benchmark Performance

Kimi K3 ranks fourth among all frontier models on the Artificial Analysis Intelligence Index, behind Claude Fable 5 and GPT-5.6 Sol, and ahead of Claude Opus 4.8.

The coding numbers are K3's headline claim. DeepSWE and FrontierSWE evaluate real-world software engineering performance — resolving open GitHub issues, writing production code — and 67.5% and 81.2% respectively place K3 among the strongest coding models available.

The LMArena Frontend Code Arena result — debuting at #1 with an Elo of 1679 — reflects user preference specifically for UI and frontend code generation. For teams building design-to-code pipelines or generating React/Vue components from screenshots, this is a meaningful signal.

On GPQA Diamond, K3 lands at approximately 88%, in the same band as most other frontier models. At the top of the scientific reasoning scale, differences between frontier models compress.

Pricing

Kimi K3 uses flat pricing across the full context window with no long-context surcharge.

The output token price is high relative to other open-weight models. At $15.00 per million, K3 costs 3.4× more per output token than GLM 5.2 ($4.40/M) and 5× more than Kimi K2.5 ($3.00/M).

The cache hit discount is significant (-90% from cache miss rate) for applications with repeated system prompts or large static context — retrieval systems, long-document Q&A with a fixed corpus, or coding agents with a codebase loaded into context. At $0.30/M on cache hits, the effective cost of repeated-context calls drops substantially.

For most developer use cases — 30K input + 10K output — a single K3 call costs roughly $0.24. A comparable GLM 5.2 call costs $0.086. At scale, that 2.8× per-call difference compounds significantly.

How to Access Kimi K3

Via API

Kimi K3's API is OpenAI SDK-compatible. The base URL is https://api.moonshot.ai/v1 and the model ID is kimi-k3.

from openai import OpenAI

client = OpenAI(
    api_key="your_moonshot_api_key",
    base_url="https://api.moonshot.ai/v1"
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Explain the trade-offs between MoE and dense transformer architectures."}
    ],
    max_completion_tokens=2048
)

print(response.choices[0].message.content)

A few API-specific notes:

• Reasoning is always on. K3 has no reasoning_effort setting — it runs at max reasoning by default, which contributes to the higher output token count and price.
• Omit temperature, top_p, and seed. K3 uses fixed sampling; passing these parameters may cause errors.
• Images require base64 or file upload. Public image URLs are not supported. Either encode images as base64 or upload via the Kimi file API and reference by ms:// file ID.
• Set max_completion_tokens explicitly. The default is 131,072 — a 128K token answer budget. For most tasks, a smaller limit saves cost significantly.

Via Chat Interface

K3 is available immediately at kimi.com and in Kimi Work and Kimi Code without an API key. The web interface exposes the full reasoning capability including the thinking trace.

Context Window and Image Input

1M context: K3 handles 1,048,576 tokens in a single request — roughly 1,500 pages of text. This is the same capacity as GLM 5.2 and significantly larger than GPT-4o's 128K or many other frontier APIs. For large codebase analysis, full legal document review, or long research paper processing, the 1M window removes the need for chunking.

Image input: K3 accepts images alongside text. This enables:

• Screenshot-to-code (describe a UI or mockup, generate implementation)
• Diagram and architecture chart understanding
• PDF and document processing with embedded images
• Visual debugging (pass a screenshot of an error state)

Images are passed as base64 or referenced via Kimi's file API. The model processes image and text in the same context window.

Open Source Status

Kimi K3's weights are scheduled for open release on July 27, 2026 — eleven days after the API launched. The license is Modified MIT, which permits commercial use with some restrictions.

Before July 27, K3 is proprietary: accessible via API and Kimi's apps, but not downloadable. After July 27, it becomes one of the largest open-weight models available for self-hosting, fine-tuning, or local deployment.

For immediate open-weight access, GLM 5.2 weights are already on Hugging Face under a full MIT license.

When to Use Kimi K3

K3 is a strong fit when:

• Your task benefits from frontier-tier reasoning — complex coding, multi-step analysis, research synthesis
• You need image input alongside text in the same API call
• GitHub-style coding is the primary workload: K3's 67.5% on DeepSWE and 81.2% on FrontierSWE reflect genuine SWE performance
• Frontend code generation is a core use case — K3's #1 LMArena Frontend Code Arena ranking reflects strong user preference for UI tasks
• You're willing to pay for the highest benchmark score currently available in open-weight form

K3 is not the best fit when:

• Your workflow is text-only and cost-sensitive at scale — GLM 5.2 at $4.40/M output is 71% cheaper
• Response speed matters for interactive users — GLM 5.2 generates 158 t/s versus K3's 62 t/s
• You need open weights now and can't wait for July 27
• The task is scientific reasoning or logic — at that level, GLM 5.2 and K3 land in the same band (≈88% GPQA Diamond) at very different prices

Kimi K3 vs GLM 5.2: The Short Version

K3 benchmarks higher and supports images. GLM 5.2 is 3.4× cheaper on output, 2.5× faster, and MIT-licensed today. For a full comparison with pricing examples, see GLM 5.2 vs Kimi K3.

Frequently Asked Questions

What is Kimi K3?

Kimi K3 is a 2.8-trillion-parameter reasoning model from Moonshot AI, released July 16, 2026. It's a Mixture-of-Experts model with a 1-million-token context window, native image input, and always-on extended reasoning. It scores 57 on the Artificial Analysis Intelligence Index — comparable to Claude Opus 4.8 — making it the highest-scoring open-weight model released as of July 2026.

Is Kimi K3 open source?

Kimi K3's API launched July 16, 2026. Open weights are scheduled for July 27, 2026, under a Modified MIT License permitting commercial use.

How many parameters does Kimi K3 have?

2.8 trillion total parameters. However, K3 is a Mixture-of-Experts model — only 16 of its 896 experts (roughly 1.8%) are activated for any given token, so the active compute is far less than the headline parameter count suggests.

How much does Kimi K3 cost?

$3.00 per million input tokens (cache miss), $0.30 per million input tokens (cache hit), and $15.00 per million output tokens. No long-context surcharge.

Does Kimi K3 support images?

Yes. K3 accepts text and image input (pass images as base64 or Kimi file IDs). It cannot generate images — output is text only.

What model ID do I use for Kimi K3's API?

The model ID is kimi-k3, called at https://api.moonshot.ai/v1. The API is OpenAI SDK-compatible.

How does Kimi K3 compare to GLM 5.2?

Kimi K3 scores 6 points higher on the AA Intelligence Index (57 vs 51) and supports image input. GLM 5.2 is 3.4× cheaper per output token ($4.40 vs $15.00/M) and generates responses 2.5× faster (158 vs 62 t/s). See the full comparison for pricing examples and benchmark breakdown.

Is Kimi K3 the same as Kimi K2?

No. Kimi K3 is the successor to the K2 series. Key differences: K3 has 2.8T parameters (vs K2.6's smaller scale), scores 57 on the AA Intelligence Index vs K2.6's lower score, and adds always-on reasoning. K3's pricing is also substantially higher than K2.x.

Bottom Line

Kimi K3 is the most capable open-weight reasoning model available as of July 2026 — 2.8T parameters, frontier-tier benchmark scores, 1M context, and image input in a single API call. It matches Claude Opus 4.8 on independent benchmarks and ranks first on LMArena's Frontend Code Arena.

The constraint is price. At $15.00/M on output tokens, K3 is expensive relative to other open-weight models. For text-only, cost-sensitive, or speed-sensitive workloads, GLM 5.2 delivers strong benchmark performance at 71% lower output cost and 2.5× faster throughput.

For developers building image-aware pipelines, high-stakes coding agents, or applications where benchmark performance justifies premium pricing: K3 is the current benchmark leader in the open-weight category. For everything else, evaluate whether the intelligence premium clears the cost bar.

Try Kimi K3 — no setup needed: glm5.app/chat?model=kimi-k3. Or compare side-by-side with GLM 5.2: glm5.app/chat.

Sources

• Kimi K3 Tech Blog: Open Frontier Intelligence — Moonshot AI (official model announcement, architecture details)
• Kimi K3 — Intelligence, Performance & Price Analysis — Artificial Analysis (Intelligence Index: 57, speed: 62 t/s, pricing data)
• Kimi K3 Quickstart — Kimi API Platform (API parameters, model ID, reasoning settings)
• China's 2.8T Kimi K3 Beats Claude Fable 5 in Frontend Code Arena — Tom's Hardware (LMArena Elo, parameter count, open weights timeline)
• Kimi K3 Is Live: Pricing, Benchmarks, and the Wait for Public Weights — Trilogy AI (pricing structure, benchmark breakdown)