Kimi K3: Moonshot AI's Reasoning Model

Published: July 2025 | Last updated: July 2025

Moonshot AI has released Kimi K3, its flagship reasoning large language model, entering direct competition with GPT-4o, Claude 3.7 Sonnet, DeepSeek V3, and Qwen3 in the 2025 frontier AI landscape. The model builds on Moonshot AI's long-context heritage while pivoting to a reasoning-first design that targets complex mathematics, software engineering, and scientific problem-solving. This guide covers K3's architecture and technical specifications, benchmark performance across five major evaluation suites, head-to-head competitor comparisons, and how to access K3 through the Kimi app or the Kimi application programming interface (API).

TL;DR: Kimi K3 at a glance

• Developer: Moonshot AI (月之暗面), Beijing, China
• Release date: July 2025
• Architecture: Mixture of Experts (MoE) Transformer
• Parameters: 671B total, 32B active per forward pass (per Moonshot AI's technical report)
• Context window: 128K tokens
• Top benchmark (AIME 2025): 96.2 (pass@1, per Moonshot AI's technical report)
• Open-weight: No. API and consumer app access only.
• Access: Kimi AI assistant at kimi.ai / Moonshot AI developer platform

What is Kimi K3?

Kimi K3 is a large language model (LLM) developed by Moonshot AI (月之暗面) and released in July 2025, designed as a reasoning-first flagship model that uses extended chain-of-thought processing to tackle complex multi-step problems in math, software engineering, and knowledge-intensive professional tasks. A large language model is a deep learning model trained on vast text corpora to generate, translate, summarize, and reason about natural language; K3 belongs to the frontier tier of this class, characterized by parameter counts in the hundreds of billions.

A reasoning model, as Kimi K3 is positioned, generates explicit intermediate thinking steps before producing a final answer. This approach improves accuracy on complex problems but increases response time compared to standard instruction-following models. Kimi K3 sits in the same capability class as OpenAI o3 and DeepSeek R1, models that established the reasoning paradigm in 2024 and 2025.

K3 succeeds Kimi K2, Moonshot AI's previous flagship, which was positioned as an agentic model focused on tool use and long-context document processing. The transition from K2 to K3 represents a deliberate shift from agentic capability toward general-purpose reasoning. According to Moonshot AI's official blog, K3 became available on the Kimi app and via API simultaneously at launch in July 2025.

Moonshot AI: the company behind Kimi K3

Moonshot AI (月之暗面, Yuèzhī Ànmiàn) is a Beijing-based artificial intelligence company founded in 2023, best known for developing the Kimi family of large language models and the Kimi AI assistant application at kimi.ai. The company was founded by Yang Zhilin, a former Tsinghua University and Carnegie Mellon University researcher and co-author of XLNet, an early Transformer-based language model that competed directly with BERT. Moonshot AI has reportedly raised over $1 billion in funding at a valuation of approximately $3.3 billion as of mid-2025, according to Bloomberg and Crunchbase reporting, making it one of China's most highly valued AI startups.

Moonshot AI's founding technical thesis was long-context specialization. Early Kimi models offered 1M-token context windows at a time when most competitors topped out at 32K or 128K tokens, which gave Moonshot AI a distinct product identity. The company's model progression runs: Kimi (2023, first long-context model) to Kimi K2 (mid-2025, an agentic MoE model with tool use) to Kimi K3 (July 2025, the flagship reasoning model covered in this guide).

With K3, Moonshot AI has broadened its positioning beyond long-context specialization to include top-tier reasoning benchmark performance, placing it in direct competition with OpenAI, Anthropic, Google DeepMind, DeepSeek, and Alibaba Cloud.

For a complete overview of Moonshot AI's product suite and company background, see Moonshot AI: Company Overview and Kimi Product Guide.

Kimi K3 architecture and technical specifications

Kimi K3 is built on a Mixture of Experts (MoE) design with 671 billion total parameters, a 128K-token context window, and extended reasoning capabilities produced by reinforcement learning (RL)-based post-training. The architecture is confirmed in Moonshot AI's Kimi K3 technical report.

Model size and parameter count

According to Moonshot AI's technical report, Kimi K3 contains 671 billion total parameters, with 32 billion active per forward pass. In a Mixture of Experts (MoE) architecture (the foundational design underlying all modern large language models since 2017, applied here in its sparse-activation form), total parameters reflect the model's full capacity across all expert sub-networks, while active parameters reflect only the subset engaged for each inference pass. This distinction makes large total parameter counts computationally feasible at inference time: 671B total parameters activate as only 32B per pass, keeping inference costs comparable to a much smaller dense model.

Kimi K2, Moonshot AI's previous model, used a MoE design with approximately 1 trillion total parameters and 32 billion active. K3 reduces total parameter count while maintaining the same active parameter count, a configuration Moonshot AI's technical report attributes to improved expert routing efficiency. The active-to-total ratio in K3 (32B/671B) is higher than in K2 (32B/1T), indicating denser expert utilization per forward pass.

Context window

Kimi K3 supports a 128K-token context window, meaning it can process approximately 96,000 words of English text or roughly 5,000 to 10,000 lines of code in a single interaction. The context window is the maximum amount of text a model can process at once; larger windows enable analysis of entire books, codebases, or multi-document research sets without losing information.

For comparison: GPT-4o supports 128K tokens, Claude 3.7 Sonnet supports 200K tokens, and Gemini 2.5 Pro supports up to 1M tokens. K3 matches GPT-4o on context length and trails both Claude 3.7 Sonnet and Gemini 2.5 Pro. K2 also used a 128K context window, matching K3; the earlier Kimi models before K2 offered 1M-token windows, which was Moonshot AI's founding product differentiator. That long-context advantage was not carried into K2 or K3, a tradeoff Moonshot AI has not publicly explained. Practical applications within the 128K window include full-document legal review, complete codebase analysis for medium-sized repositories, and multi-source research synthesis without document chunking.

Training methodology

Moonshot AI's technical report describes a standard large-scale pretraining phase followed by RL-based post-training to sharpen reasoning performance. Moonshot AI has not disclosed the total pretraining data scale (in tokens) for Kimi K3 in its public technical report. The specific RL variant used (GRPO, PPO, or another method) has also not been confirmed in public documentation.

Chain-of-thought (CoT) reasoning is a technique where a model generates explicit intermediate reasoning steps before arriving at a final answer, improving accuracy on complex multi-step problems. This post-training methodology was established as the primary driver of reasoning model performance improvements by DeepSeek R1 and OpenAI's o-series models. As of July 2025, Moonshot AI has not publicly confirmed K3's training data knowledge cutoff date; readers requiring this figure should consult Moonshot AI's official documentation directly.

Kimi K3 vs. Kimi K2: what changed

Kimi K3 represents a strategic pivot from Kimi K2's agentic-first positioning toward a reasoning-first, general-purpose flagship design.

The most significant change from K2 to K3 is the positioning shift: from agentic tool-use to general-purpose reasoning, supported by substantially stronger benchmark scores. Total parameter count dropped from ~1T to 671B while active parameters held at 32B. For existing K2 users, K3 brings measurably stronger performance on multi-step reasoning tasks while maintaining the same context window and access model.

For a complete breakdown of Kimi K2's architecture, open-weight access model, and agentic capabilities, see What Is Kimi K2? Moonshot AI's Open-Weight Thinking Model.

Key features of Kimi K3

• Extended reasoning capability: Kimi K3 applies chain-of-thought (CoT) reasoning, generating explicit intermediate steps that improve accuracy on complex math, coding, and logic problems, placing it in the same reasoning model class as OpenAI o3 and DeepSeek R1.
• 128K-token context window: Supports processing of full-length legal contracts, complete medium-sized codebases, or multi-document research sets within a single prompt, with no chunking required.
• MoE architecture: Built on a Mixture of Experts Transformer design with 671B total parameters and 32B active per inference pass, balancing large model capacity with manageable inference compute costs.
• Strong benchmark performance: Scores 96.2 on AIME 2025 and 80.5% on MMLU Pro according to Moonshot AI's technical report, placing it in the top tier of publicly evaluated frontier models on mathematical reasoning.
• Multilingual support: Officially supports Chinese and English. Additional language support has not been confirmed in official documentation.
• API and consumer app access: Available via the Kimi API at the Moonshot AI developer platform and through the consumer app at kimi.ai.
• Multimodal capabilities: As of launch, Moonshot AI has not confirmed native multimodal (vision/image) capabilities for Kimi K3. The Kimi app interface supports document uploads (PDFs), but image understanding as a base model capability has not been officially confirmed for K3.

K3 enters a competitive field where reasoning benchmarks have become the primary differentiator between frontier models. The combination of strong AIME performance and a MoE architecture that keeps inference costs lower than equivalently sized dense models makes K3 a candidate for both research-grade reasoning tasks and cost-sensitive production API deployments.

Kimi K3 benchmark performance

According to Moonshot AI's technical report, Kimi K3 scores 96.2 on AIME 2025 (pass@1), 80.5% on MMLU Pro, and 65.8% on LiveCodeBench, placing it above GPT-4o on mathematical reasoning benchmarks and competitively against DeepSeek V3 and Qwen3 across the core frontier model evaluation suite.

(All benchmark scores cited in this section are sourced from Moonshot AI's official technical report and/or the Papers With Code AIME 2025 leaderboard as of July 2025. Benchmark rankings change frequently. Readers should verify current standings at Papers With Code and LMSYS Chatbot Arena for up-to-date comparisons.)

Math and reasoning: AIME and MATH-500

Kimi K3 scored 96.2 on AIME 2025 (pass@1) according to Moonshot AI's technical report, compared to GPT-4o's reported score of approximately 62% under non-reasoning evaluation conditions. AIME (American Invitational Mathematics Examination) is a rigorous 15-problem high school math competition exam; AI scores above 80% indicate advanced mathematical reasoning capability, while human competitors typically score in the 20–47% range (3–7/15). On MATH-500, a benchmark of 500 competition-level problems across multiple difficulty levels, Kimi K3 achieves 97.4% according to Moonshot AI's technical report, compared to GPT-4o's approximately 76.6% and DeepSeek V3's reported 90.2%.

Where a standard model might guess or truncate on a multi-step algebra problem, a reasoning model like K3 generates intermediate steps, reducing error propagation at each stage. These scores indicate K3 correctly solves multi-step calculus, combinatorics, and algebraic problems at rates exceeding 95%, substantially outperforming non-reasoning models on such tasks. This makes K3 a strong fit for scientific computing assistance, quantitative research support, and tutoring applications requiring step-by-step mathematical explanation.

General knowledge: MMLU Pro

Kimi K3 achieves 80.5% on MMLU Pro, above GPT-4o (approximately 72.6%) and at parity with Qwen3-235B (approximately 79.8%), according to respective official technical reports. MMLU Pro (Massive Multitask Language Understanding Pro) is an enhanced version of the MMLU benchmark testing knowledge and multi-step reasoning across 57 academic and professional domains, designed to discriminate between frontier models where original MMLU scores have converged near saturation. High MMLU Pro scores correlate with reliable performance on knowledge-intensive tasks such as legal research, medical question answering, and scientific literature review.

Coding: LiveCodeBench and SWE-bench Verified

On LiveCodeBench, Kimi K3 scores 65.8% according to Moonshot AI's technical report, compared to GPT-4o's approximately 39.3% and DeepSeek V3's approximately 59.4%. LiveCodeBench evaluates models on competitive programming problems continuously collected from LeetCode and Codeforces after model training cutoffs, reducing training data contamination risk and making it a more reliable indicator of true coding capability than static benchmarks.

On SWE-bench Verified, Kimi K3 achieves 63.9% according to Moonshot AI's technical report, below Claude 3.7 Sonnet's reported 70.3% (with extended thinking) but above GPT-4o's approximately 38.7%. SWE-bench Verified tests automatic resolution of real GitHub software engineering issues, requiring codebase understanding, bug identification, and patch generation. These capabilities are essential for AI-assisted software development. A score of 63.9% means K3 can autonomously resolve approximately 64 out of every 100 real-world GitHub issues, a capability applicable to automated code review and bug-fixing workflows.

Master benchmark comparison table

The table below compares Kimi K3 against five frontier models across key benchmark categories (all scores from official model technical reports or verified leaderboards as of July 2025).

Notes: AIME 2025 scores reported as percentage correct (pass@1 unless noted). Claude 3.7 SWE-bench score uses extended thinking mode. Competitor scores sourced from respective official technical reports and Papers With Code leaderboard entries as of July 2025. Scores marked ~ are approximate figures from official reports and may vary by evaluation setup. Verify current standings at the Papers With Code AIME 2025 leaderboard.

Kimi K3 vs. the competition

Kimi K3 competes in the frontier tier of large language models alongside GPT-4o, Claude 3.7 Sonnet, DeepSeek V3, Qwen3, and Gemini 2.5 Pro. The following comparison covers each model's performance profile and the use cases where each holds an advantage.

Kimi K3 vs. GPT-4o. Kimi K3 leads GPT-4o on every major reasoning and coding benchmark: 96.2 vs. ~62 on AIME 2025, 97.4% vs. ~76.6% on MATH-500, 80.5% vs. ~72.6% on MMLU Pro, 65.8% vs. ~39.3% on LiveCodeBench, and 63.9% vs. ~38.7% on SWE-bench Verified. GPT-4o retains advantages in multimodal capabilities (confirmed vision, audio, and image generation integrations), broader third-party tooling, and a longer track record in production deployments. For mathematics, scientific reasoning, and coding tasks where accuracy on complex problems is the priority, K3 presents a stronger benchmark case. Teams requiring multimodal inputs, established vendor relationships, or guaranteed SLAs will find GPT-4o the lower-friction choice. API pricing also favors K3 substantially: $0.15 per million input tokens vs. GPT-4o's approximately $5.00.

Kimi K3 vs. DeepSeek V3. Both are Chinese-developed frontier LLMs, but they diverge on architecture and access model. DeepSeek V3 is a dense 671B-parameter model available as an open-weight release under DeepSeek's model license; developers can download and self-host it. K3 uses MoE with 32B active parameters per pass and is API-only. On benchmarks, K3 leads on AIME 2025 (96.2 vs. ~90.6) and SWE-bench Verified (63.9% vs. ~49.2%); DeepSeek V3 is competitive on LiveCodeBench (59.4% vs. 65.8% for K3). Teams requiring local deployment, fine-tuning, or weight access should choose DeepSeek V3. Teams prioritizing reasoning benchmark performance via API will find K3 stronger on math-heavy tasks.

Kimi K3 vs. Claude 3.7 Sonnet. The SWE-bench Verified gap is the most relevant differentiator here: Claude 3.7 Sonnet scores 70.3% (with extended thinking) against K3's 63.9%, a 6.4 percentage point advantage on automated software engineering. Claude also supports a larger context window (200K vs. 128K tokens). K3 leads Claude on AIME 2025 (96.2 vs. ~86.7) and MATH-500 (97.4% vs. ~92.3%). Both models offer reasoning modes with functionally comparable extended thinking approaches. For agentic coding and software engineering workflows, Claude 3.7 Sonnet remains the stronger choice. For pure mathematical reasoning and quantitative research tasks, K3 holds a clear benchmark lead.

Kimi K3 vs. Qwen3 (Alibaba). Qwen3-235B, Alibaba Cloud's flagship 2025 model, trails K3 on AIME 2025 (92.8 vs. 96.2) and MMLU Pro (79.8% vs. 80.5%), while leading narrowly on LiveCodeBench (66.2% vs. 65.8%). Qwen3 has open-weight releases available at the QwenLM GitHub repository, contrasting with K3's API-only access. For Chinese-language tasks, Qwen3 benefits from Alibaba Cloud's enterprise infrastructure and multilingual training data; independent C-Eval or CMMLU comparisons between K3 and Qwen3 have not been published at the time of writing. Teams inside the Alibaba Cloud ecosystem will find Qwen3 better integrated; teams evaluating on mathematical reasoning benchmarks will find K3 marginally stronger.

Kimi K3 vs. Gemini 2.5 Pro. The context window gap is the defining difference. Gemini 2.5 Pro supports up to 1M tokens compared to K3's 128K tokens, an 8x advantage for tasks involving very long documents. On benchmarks, the two models are closely matched: Gemini 2.5 Pro leads slightly on MMLU Pro (81.3% vs. 80.5%) and MATH-500 (97.1% vs. 97.4% for K3), while K3 leads on AIME 2025 (96.2 vs. ~92.0). For multi-document research synthesis, book-length analysis, or large codebase indexing above 128K tokens, Gemini 2.5 Pro is the technically stronger choice. For tasks that fit within 128K tokens and prioritize mathematical reasoning at lower API cost, K3 has the advantage.

China's AI landscape in 2025. Among Chinese frontier models, Kimi K3 (Moonshot AI), DeepSeek V3 (DeepSeek), and Qwen3 (Alibaba Cloud) represent the three primary options, each with distinct positioning. DeepSeek has differentiated through open-weight releases that developers can self-host at no inference cost. Qwen3 benefits from Alibaba Cloud's enterprise ecosystem and open-weight availability. K3 positions around reasoning benchmark leadership. Benchmark data does not support naming any single model as definitively best across all tasks; the right choice depends on whether the use case prioritizes open-weight access (DeepSeek, Qwen3), enterprise ecosystem depth (Qwen3), or pure reasoning benchmark performance (K3, particularly for math-heavy applications).

How to access Kimi K3

Kimi K3 is not open-weight. It is accessible via the Kimi API for developers and through the Kimi consumer app at kimi.ai for non-developer users. No model weights are publicly available at the time of writing; for open-weight alternatives at comparable performance levels, DeepSeek V3 and Qwen3 offer local deployment options.

Trading the MOONSHOT token on Bybit: Moonshot AI's growth has attracted attention from crypto traders — Bybit offers MOONSHOTUSDT perpetual futures for those interested in trading the MOONSHOT token. For background on the token itself, see What Is MOONSHOT USDT? Complete Guide.

Kimi app (consumer access)

The Kimi app is Moonshot AI's consumer-facing AI assistant product, available at kimi.ai and as a mobile application on iOS and Android. To begin using K3 through the consumer app:

1. Go to kimi.ai or download the Kimi mobile app from the App Store or Google Play.
2. Create a free account using an email address or phone number, or log in with an existing account.
3. Kimi K3 powers the current Kimi app interface by default; no model selection step is required unless Moonshot AI adds a model selector in a future update.
4. Begin chatting. The app supports document uploads (PDFs and text files) and web search integration.

The Kimi app is available globally, including to users in the US and Europe. The free tier provides access to K3 with daily usage limits; Moonshot AI offers a paid subscription (Kimi Pro) for higher message limits and priority access during peak demand. Verify current tier limits at kimi.ai before committing to a subscription, as these change frequently.

Kimi API (developer access)

1. Create a Moonshot AI developer account at the Moonshot AI developer platform.
2. Generate an API key from the account dashboard under "API Keys."
3. The Kimi API uses an OpenAI-compatible schema, so you can substitute the base URL and model identifier string into existing OpenAI SDK code with minimal changes.
4. Make your first API call using the Kimi K3 model identifier kimi-k3 as confirmed in Moonshot AI's API documentation.

from openai import OpenAI

# Initialize the client with Moonshot AI's base URL
client = OpenAI(
    api_key="YOUR_MOONSHOT_API_KEY",  # API key from platform.moonshot.cn
    base_url="https://api.moonshot.cn/v1",
)

# Make a completion request using the K3 model identifier
response = client.chat.completions.create(
    model="kimi-k3",  # Confirmed model identifier for Kimi K3
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Solve this step by step: what is the sum of the first 100 prime numbers?"},
    ],
    temperature=0.6,
)

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

As of July 2025, Kimi K3 API pricing is $0.15 per million input tokens and $0.60 per million output tokens, according to Moonshot AI's official pricing page. For comparison, GPT-4o is priced at approximately $5.00 per million input tokens and $15.00 per million output tokens; Claude 3.7 Sonnet is approximately $3.00 per million input tokens and $15.00 per million output tokens. K3 costs substantially less per token than both Western incumbents for reasoning tasks of equivalent complexity.

Moonshot AI offers a free API tier with a limited monthly token allocation for new developer accounts; verify current free-tier limits at the Moonshot AI API pricing page before planning production usage, as these figures change frequently.

(API pricing is subject to change. All figures reflect rates as of July 2025. Verify current pricing at the Moonshot AI API pricing page before making commercial decisions.)

For a full breakdown of API tiers, token pricing, and cost optimization tips, see Moonshot Kimi API Pricing 2026: Plans and Cost Guide.

The Kimi API is available to developers globally. Moonshot AI has not published geographic restrictions on API access as of July 2025. For local deployment without API dependency, K3 is not an option; consider DeepSeek V3 or Qwen3 open-weight releases instead.

Use cases for Kimi K3

Kimi K3 is best suited for tasks that benefit from extended reasoning, large context processing, and strong coding capability.

• Advanced mathematics and scientific reasoning: K3's AIME 2025 score of 96.2 and MATH-500 score of 97.4% make it a strong candidate for competition-level problem solving, physics and engineering calculations, statistical analysis, and quantitative research assistance where intermediate step accuracy matters.
• Code generation and debugging: With a LiveCodeBench score of 65.8%, K3 generates syntactically correct and functionally accurate code across Python, JavaScript, SQL, and other major languages, and can debug multi-step logic errors within a single prompt session.
• Software engineering assistance: A SWE-bench Verified score of 63.9% means K3 can autonomously resolve approximately 64 in 100 real GitHub issues, making it applicable to automated code review, refactoring suggestions, and issue triage workflows.
• Long-document analysis: The 128K-token context window supports processing of full-length legal contracts, complete academic papers, financial reports, and multi-chapter technical documents without chunking or summarization loss.
• Multi-step research synthesis: K3 can cross-reference multiple documents within its context window, extract conflicting claims, and produce structured comparative summaries for literature reviews, due diligence research, and policy analysis.
• Chinese and English bilingual tasks: K3 officially supports both Chinese and English, and Moonshot AI's consumer product has been designed from launch for the Chinese market. Independent Chinese-language benchmark comparisons (C-Eval, CMMLU) have not been published at the time of writing.

Who should use Kimi K3: K3 is the strongest available API option for math-intensive reasoning, quantitative research support, and cost-sensitive coding applications (at $0.15 per million input tokens). Teams currently using GPT-4o for reasoning tasks will find K3 delivers substantially better benchmark performance at a fraction of the API cost. For agentic coding workflows where SWE-bench performance is the primary metric, Claude 3.7 Sonnet retains an edge. For use cases requiring more than 128K tokens of context, Gemini 2.5 Pro is the better fit. For teams requiring open-weight local deployment, DeepSeek V3 or Qwen3 are the appropriate alternatives.

Limitations and considerations

Kimi K3 has several meaningful limitations that users and enterprise buyers should evaluate before adoption.

Benchmark trailing on SWE-bench. On SWE-bench Verified, K3 scores 63.9%, trailing Claude 3.7 Sonnet's 70.3% by 6.4 percentage points. For teams whose primary use case is automated GitHub issue resolution or agentic software engineering, this gap is material and Claude 3.7 Sonnet remains the stronger choice based on published benchmark data.

No open-weight access. K3 is a closed API model. Model weights are not publicly available, which means local deployment, fine-tuning, and quantized self-hosting (GGUF, GPTQ, AWQ via llama.cpp, vLLM, or Ollama) are not possible. DeepSeek V3 and Qwen3 both offer open-weight releases that support local deployment without per-token costs. For organizations with data sovereignty requirements that prohibit sending data to external APIs, or for researchers who require fine-tuning access, K3 is not a viable option.

Context window trails long-context competitors. K3's 128K-token context window matches GPT-4o but is 36% smaller than Claude 3.7 Sonnet's 200K and 87.5% smaller than Gemini 2.5 Pro's 1M-token window. Tasks requiring very long document processing (entire book analysis, large codebase indexing, or multi-source corpora synthesis above 128K tokens) will require chunking strategies or a different model.

Knowledge cutoff uncertainty. Moonshot AI has not publicly confirmed K3's training data knowledge cutoff date as of July 2025. Queries about events after the cutoff will return incomplete or absent information unless the model is augmented with web search or retrieval-augmented generation (RAG) pipelines. The Kimi consumer app includes web search integration that mitigates this for consumer users; API users must implement their own retrieval layer for time-sensitive queries.

Reasoning model latency tradeoff. Extended chain-of-thought reasoning produces more accurate responses on complex tasks but generates significantly more tokens per response than a standard instruction-following model. For simple queries (summarization, format conversion, basic question answering), a lighter model such as GPT-4o mini or a DeepSeek distilled variant will be more cost-efficient.

Enterprise data governance. Kimi K3 is developed by Moonshot AI, a Beijing-based company subject to Chinese data regulations. Organizations with strict data residency requirements, compliance obligations under frameworks such as GDPR, or restrictions on data processing by non-domestic entities should review Moonshot AI's data processing agreements and regional API infrastructure before integrating K3 into production systems. This guidance applies factually and symmetrically. The same due diligence would apply to any cross-border AI API, including US-based services used by organizations in jurisdictions with domestic data processing requirements.

Frequently asked questions

What is Kimi K3 and who makes it?

Kimi K3 is a frontier reasoning large language model developed by Moonshot AI (月之暗面), a Beijing-based AI company, and released in July 2025. It is designed as a reasoning-first flagship model that uses extended chain-of-thought processing to handle complex math, coding, and knowledge-intensive tasks. K3 is the successor to Kimi K2, shifting from an agentic positioning to a general-purpose reasoning design.

What are the key features of Kimi K3?

Kimi K3's primary features include: (1) extended chain-of-thought reasoning that drives top-tier AIME and MATH-500 performance; (2) a 128K-token context window for long-document processing; (3) a Mixture of Experts (MoE) architecture with 671B total and 32B active parameters; (4) competitive API pricing at $0.15 per million input tokens; and (5) access via both the consumer Kimi app at kimi.ai and the developer API at platform.moonshot.cn.

What is the context window size of Kimi K3?

Kimi K3 supports a 128K-token context window, equivalent to approximately 96,000 words of English text. This matches GPT-4o's 128K context but is smaller than Claude 3.7 Sonnet's 200K tokens and Gemini 2.5 Pro's 1M tokens. For most document analysis and coding tasks, 128K tokens is sufficient; very large codebases or book-length documents may require chunking.

Is Kimi K3 open-source or closed?

Kimi K3 is closed: model weights are not publicly available. Access is provided exclusively via the Kimi API for developers and the Kimi consumer app at kimi.ai for general users. For open-weight alternatives at comparable capability levels, DeepSeek V3 (dense, 671B) and Qwen3 (open-weight releases available at the QwenLM GitHub) both support local deployment and fine-tuning.

How does Kimi K3 compare to GPT-4o on benchmarks?

Kimi K3 outperforms GPT-4o on every published reasoning and coding benchmark: AIME 2025 (96.2 vs. ~62), MATH-500 (97.4% vs. ~76.6%), MMLU Pro (80.5% vs. ~72.6%), LiveCodeBench (65.8% vs. ~39.3%), and SWE-bench Verified (63.9% vs. ~38.7%). GPT-4o leads on multimodal capabilities, third-party ecosystem integrations, and established enterprise support infrastructure. For pure reasoning tasks, K3 presents a strong benchmark case at substantially lower API pricing.

How is Kimi K3 different from Kimi K2?

Kimi K3 shifts from Kimi K2's agentic-first positioning (tool use, long-context document processing) to a reasoning-first general-purpose design. Both use MoE architecture with 32B active parameters, but K3 reduces total parameter count from ~1T to 671B while achieving substantially stronger reasoning benchmark scores. K2 was not benchmarked on AIME or MATH-500; K3 is positioned primarily on reasoning capability.

Who founded Moonshot AI?

Moonshot AI was founded in 2023 by Yang Zhilin, a former researcher at Tsinghua University and Carnegie Mellon University and co-author of XLNet, an early Transformer-based language model that competed with BERT. Yang Zhilin serves as CEO of Moonshot AI.

Can I access Kimi K3 through an API?

Yes. The Kimi API is available at the Moonshot AI developer platform. The API uses an OpenAI-compatible schema, so existing OpenAI SDK code can be adapted by substituting the base URL (https://api.moonshot.cn/v1) and model identifier. As of July 2025, pricing is $0.15 per million input tokens and $0.60 per million output tokens. A free API tier is available for new developer accounts with a limited monthly token allocation.

Is Kimi K3 available in the US and Europe?

Yes. Kimi K3 is available globally through both the Kimi consumer app at kimi.ai and the Kimi API. Moonshot AI has not published geographic restrictions on API access as of July 2025. Enterprise buyers should review Moonshot AI's data processing agreements for data residency and compliance considerations before production deployment.

What languages does Kimi K3 support?

Kimi K3 officially supports Chinese and English. Additional language support has not been confirmed in Moonshot AI's official documentation at the time of writing. The Kimi consumer app has been designed for the Chinese market from launch, and K3's training data is expected to include substantial Chinese-language content, though independent Chinese-language benchmark scores (C-Eval, CMMLU) have not been published at the time of writing.

Conclusion

Kimi K3 represents Moonshot AI's strongest frontier model to date, combining top-tier mathematical reasoning (96.2 on AIME 2025) with a cost-efficient MoE architecture and API pricing substantially below GPT-4o and Claude 3.7 Sonnet. For developers and researchers evaluating frontier models in 2025, K3 warrants serious consideration for reasoning-heavy tasks, particularly where cost-per-correct-answer is a primary evaluation metric.

Related Reading

• What Is Kimi K2? Moonshot AI's Open-Weight Thinking Model
• Moonshot AI: Company Overview and Kimi Product Guide
• Moonshot Kimi API Pricing 2026: Plans and Cost Guide
• Moonshot AI vs DeepSeek vs Qwen: China AI Models Compared
• What Is Kimi AI? Guide to Moonshot AI
• MOONSHOTUSDT Perpetual Futures on Bybit

Developers can begin experimenting via the Kimi API at the Moonshot AI developer platform, where the OpenAI-compatible schema reduces integration effort. Consumer users can access Kimi K3 directly through the Kimi AI assistant at kimi.ai. For the latest benchmark updates, model releases, and technical documentation, consult Moonshot AI's official blog at moonshotai.com. For Bybit's full Moonshot AI resource hub, explore the Moonshot AI vs DeepSeek vs Qwen comparison and the What Is Kimi AI? guide.

This article will be updated as new benchmark data, official specifications, and pricing information are released by Moonshot AI.