Moonshot AI — Open-Weight Frontier Model July 16, 2026
20 free Kimi K3 prompts for coding, long-context reasoning, vision, research, and agentic workflows — built for Moonshot AI's 2.8T open-weight model. Copy directly into the Kimi app or API. No signup.
The Kimi K3 prompt generator on this page gives you 20 copy-paste prompts designed to unlock the core capabilities of Moonshot AI's largest model. Kimi K3 launched July 16–17, 2026 as the first open-weight model in the 3-trillion-parameter class — a threshold previously reached only by closed proprietary systems. With a 2.8T Mixture-of-Experts architecture, a 1M token context window, and native vision, Kimi K3 is Moonshot's answer to Claude Fable 5, GPT-5.6 Sol, and Gemini 4 — and it outperforms Claude Opus 4.8 and GPT-5.5 on coding and agentic benchmarks.
The prompts are organised by use case: coding (K3's standout strength), long-context document processing, vision tasks, research synthesis, reasoning, and agentic workflows. Each prompt is designed to leverage the unique features Moonshot built in — particularly the ability to paste entire codebases or document libraries into a single query. Paste any prompt directly into kimi.ai or your Kimi API call, fill in the bracketed placeholders, and copy.
896 expert sub-networks, 16 activated per token. Frontier intelligence at a fraction of a dense model's compute cost.
Paste an entire codebase, a year of docs, or hundreds of papers in one query. Kimi Delta Attention keeps memory costs low at long context.
Reasoning mode runs by default on every query — K3 thinks before it answers without needing a special prompt flag.
Full weights releasing July 27, 2026. Download, fine-tune, and run Kimi K3 on your own infrastructure — no per-token fees.
Optimised for Kimi K3's 1M-token window, native reasoning mode, and vision capabilities. Works with kimi.ai and the Kimi API.
Build a production-ready [component name] in [React / Vue / Svelte] with TypeScript. Requirements: - Props: [list your props with types] - Behaviour: [describe what it does] - Edge cases to handle: [empty state, loading, error] - Styling: [Tailwind CSS / CSS Modules / styled-components] Deliver: 1. The complete component file 2. A TypeScript interface for props 3. Three usage examples showing different prop combinations 4. Any accessibility (ARIA) attributes needed
Diagnose this bug systematically. Work through every plausible cause before settling on the most likely one. Error: ``` [Paste error message or stack trace] ``` Relevant code: ``` [Paste the code] ``` Environment: [OS, runtime, package versions] Already tried: [list what you've attempted] Deliver: 1. Ranked hypothesis list (most → least likely) with one-line reasoning for each 2. The fix for the top cause — exact line changes 3. Steps to take if the top fix doesn't work 4. How to prevent this class of bug in future
I am pasting our entire codebase below (up to 1M tokens). Review it and identify: 1. ARCHITECTURE ISSUES — structural problems that will compound as the codebase grows 2. SECURITY VULNERABILITIES — any code that could be exploited (injection, insecure auth, exposed secrets) 3. PERFORMANCE HOTSPOTS — functions or patterns that will not scale 4. DEAD CODE — unused exports, imports, or entire files 5. MISSING TESTS — the 5 highest-risk paths with no test coverage For each finding: file, line number, one-sentence description, and the specific fix. [Paste codebase here]
Optimise this SQL query for maximum performance. Explain every change. Database: [PostgreSQL / MySQL / SQLite] Table sizes: [e.g. orders: 8M rows, users: 600K rows] Indexes that exist: [list them] Current query: ```sql [Paste your query] ``` Deliver: 1. The rewritten query 2. A bullet list of every change and why it helps 3. Any new indexes to create (exact CREATE INDEX statement) 4. The expected performance improvement (rough order of magnitude)
Synthesise the following research papers into a structured report for a technical audience. Papers: [paste abstracts, or full text if short enough] Topic: [the shared subject] Audience: [describe their background] Report structure: 1. CONSENSUS — what all papers agree on 2. CONTRADICTIONS — where they disagree and why 3. METHODOLOGY GAPS — weaknesses in the evidence base 4. PRACTICAL TAKEAWAYS — 3 actionable conclusions for someone working in this field 5. OPEN QUESTIONS — what research still needs to be done Target length: 600–800 words.
Help me choose between these architectural approaches for the following system. System: [describe what you're building] Scale targets: [users, requests/sec, data volume] Team size: [engineers, their expertise] Options: - Option A: [describe] - Option B: [describe] - Option C: [describe, or omit if only two] Deliver: 1. A weighted trade-off matrix (latency, operational complexity, cost, scalability, team fit) 2. Your recommendation with a clear rationale 3. The strongest argument AGAINST your recommendation 4. The first proof-of-concept I should build to validate the key assumption
I am attaching a screenshot of a UI. Convert it to clean, working [React / HTML+Tailwind / Vue] code. Instructions: - Match the layout as closely as possible - Use semantic HTML elements - Make it responsive (mobile-first) - Replace any real text with realistic placeholder text - Do not use inline styles — use utility classes After the code, add: - A list of any design decisions you made where the screenshot was ambiguous - Any accessibility improvements beyond what the screenshot shows [Attach screenshot]
Execute the following multi-step task. Before each step, state what you are doing and what information you need. Pause and ask if a decision requires my input. Task: [describe the goal, e.g. 'Analyse our Q3 support tickets, identify the top 5 recurring issues, draft a proposed fix for each, and create a prioritised roadmap'] Available context: [Paste your data, documents, or starting information] After all steps, write: 1. A one-paragraph executive summary of what you found 2. The single highest-priority action I should take this week 3. What you would do next if you had more time
Prove the following statement. Show every deduction explicitly — do not skip steps. Statement: [state the proposition to prove] Allowed axioms / theorems: [list any you want to restrict to, or say 'standard undergraduate mathematics'] Preferred proof style: [direct / contradiction / induction / construction] Format: - Number each step - State the rule or theorem applied at each step in parentheses - End with a clear QED statement If the statement is false, prove why and provide the correct version.
Write a competitive landscape brief for the following market. Our product: [one sentence description] Target customer: [describe] Competitors to analyse: [list 5–6 companies] For each competitor: - Their primary value proposition (one sentence) - Estimated customer profile - One thing they do better than us - One thing we do better than them - Their apparent pricing model End with: the white space in this market — what positioning is genuinely unclaimed that we could own in the next 12 months?
Design a RESTful API for the following system. Deliver a complete OpenAPI 3.1 YAML spec. System: [describe what the API controls] Key entities: [list them — e.g. users, orders, products] Auth: [bearer token / API key / OAuth2] Special requirements: [pagination, rate limiting, webhooks, etc.] For each endpoint include: - Method and path - Description - Request body schema with validation rules - Response schemas (200 and all error codes) - One curl example Also write a short README section explaining the authentication flow.
Generate a comprehensive test suite for the following function using [Jest / Vitest / pytest]. Function: ``` [Paste the function] ``` Cover: - Happy path (normal inputs) - Edge cases (empty, null, zero, negative) - Boundary conditions (min/max values) - Invalid inputs (wrong types, malformed data) - Async behaviour (if applicable) For each test: - Use a descriptive name: what scenario + what result - Follow AAA pattern (Arrange / Act / Assert) - Add a one-line comment explaining why this case matters
I am pasting a long document below. Extract the key information into a structured summary. Document type: [contract / research report / technical spec / meeting transcript / legal filing] My role: [describe who you are and what decision you need to make] Most important questions to answer: 1. [Your first question] 2. [Your second question] 3. [Your third question] Deliver: 1. A 200-word executive summary 2. Direct answers to my three questions with page/section references 3. Any critical information I did NOT ask about but should know 4. A list of any ambiguous or missing information [Paste document]
Analyse the chart or data table in the attached image. Context: [describe what the data is measuring, the time period, and the source if known] Decision I'm making: [describe what business or technical decision this data informs] Deliver: 1. The single most important trend in this data 2. The most surprising or counterintuitive finding 3. What the data does NOT tell me that I might wrongly assume 4. One sentence recommendation based on this data 5. What additional data I should collect to be more confident [Attach chart or table image]
Audit the following code for security vulnerabilities. Check against OWASP Top 10 and common patterns for this stack. Stack: [language, framework, database] Code: ``` [Paste the code] ``` For each vulnerability found: - Severity: Critical / High / Medium / Low - CWE number (if applicable) - Description: what is wrong and how it could be exploited - Fix: the specific code change needed - Prevention: how to avoid this class of issue in future End with a one-paragraph overall security assessment.
Migrate the following code from [source language/framework] to [target language/framework]. Source code: ``` [Paste the code to migrate] ``` Migration requirements: - Preserve all existing behaviour exactly - Follow [target framework]'s idiomatic patterns and conventions - Replace deprecated or unavailable APIs with the target ecosystem's equivalents - Add TypeScript types / type hints if the target language supports them After the migrated code: 1. A diff-style summary of what changed and why 2. Any behaviour differences I should test manually 3. Dependencies to add (exact package names and versions)
Write a 1,500-word technical article for [your publication / audience]. Topic: [your topic] Primary keyword: [must appear in H1, intro, and at least twice in body] Audience expertise level: [beginner / intermediate / senior engineer] Tone: [conversational / authoritative / tutorial-style] Structure: - H1 containing the primary keyword - 120-word intro that states the reader's problem and promises a specific takeaway - 4–5 H2 sections, each with a concrete example (not hypothetical) - Code snippets where relevant - 100-word conclusion with a single clear next step Do not use filler phrases or restate the intro in the conclusion.
Design a data pipeline architecture for the following use case. Data sources: [list them — databases, APIs, file uploads, streams] Data volume: [GB/day, rows/hour, or events/second] Output consumers: [dashboards, ML models, data warehouse, APIs] Latency requirement: [real-time <1s / near-real-time <5min / batch daily] Team: [engineers and their expertise with data tools] Deliver: 1. An architecture diagram described in text (source → transform → sink) 2. Tool recommendations for each layer with one-line rationale 3. The two biggest failure modes and how to mitigate them 4. A phased rollout plan: MVP (1 week) → production-ready (1 month)
Prepare me for a technical interview at [company type / stage] for a [role title] position. My background: [2–3 sentences on your experience] Role focuses on: [key responsibilities from the job description] Interview format: [coding round / system design / behavioural / all three] Deliver: 1. The 6 most likely technical questions for this specific role 2. A strong answer framework for each — key points to hit, not a full script 3. One system design question I should expect and a structured approach to answer it 4. The 3 questions I should ask the interviewer to signal strategic thinking 5. One weakness I should avoid mentioning given this role
Write a complete Product Requirements Document for the following feature. Feature: [name] Problem: [one sentence on the pain point] Target user: [describe] Success metric: [how will you know it worked?] PRD must include: 1. PROBLEM STATEMENT — 2 paragraphs: current pain and why it matters now 2. PROPOSED SOLUTION — what we are building (not how) 3. USER STORIES — 5 stories: 'As a [user], I want [goal], so that [reason]' 4. ACCEPTANCE CRITERIA — testable conditions for each user story 5. OUT OF SCOPE — 3 things explicitly NOT in v1 6. OPEN QUESTIONS — 3 undecided items and who owns each 7. TIMELINE — phases (discovery / build / launch) in sequence, no dates
How Kimi K3 stacks up against the top frontier models on context, reasoning, openness, and pricing.
| Model | Best For | Context | Open Weight | Price (input/output) |
|---|---|---|---|---|
| Kimi K3 ✓ | Coding, long-context, open-weight self-hosting | 1M tokens | Yes (July 27) | $3 / $15 per MTok |
| Claude Fable 5 | Hardest frontier tasks, legal, full-codebase audits | 1M tokens | No | $10 / $50 per MTok |
| GPT-5.6 Sol | Broad general use, tool use, Codex integration | 256K tokens | No | $5 / $30 per MTok |
| Gemini 4 | Multimodal tasks, Google ecosystem | 1M tokens | No | $7 / $28 per MTok |
| DeepSeek V4 | Open-source reasoning, math, code | 1M tokens | Yes (MIT) | $0.27 / $1.10 per MTok |
| Grok 4.5 | Reasoning, X/Twitter integration, coding | 256K tokens | No | $2 / $6 per MTok |
Kimi K3 is Moonshot AI's flagship language model, released July 16–17, 2026. It is a Mixture-of-Experts (MoE) model with approximately 2.8 trillion total parameters — the first open-weight model in the 3-trillion-parameter class. Kimi K3 activates 16 of its 896 expert sub-networks per token using a novel architecture called Kimi Delta Attention combined with Attention Residuals, which Moonshot claims delivers 2.5× better scaling efficiency than Kimi K2. It features a 1M token context window, native vision, and a reasoning mode that runs by default on every query.
Kimi K3 is open-weight, not fully open-source. As of July 17, 2026, the model is available via the Kimi app and Kimi Code, and full model weights are scheduled for release by July 27, 2026. Open-weight means the parameter files will be downloadable and customisable by external developers and companies, but the training code and data may not be published. This places it in the same category as Meta Llama and DeepSeek — downloadable, fine-tuneable, but with some transparency limits.
Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 on coding and agentic benchmarks, and trails Claude Fable 5 and GPT-5.6 Sol on overall performance according to Moonshot's own evaluation data. Bloomberg and Tom's Hardware benchmark analysis shows K3 narrowing the gap with US frontier models significantly — it beats Claude Fable 5 specifically on the Frontend Code Arena benchmark. For coding, research synthesis, and long-context tasks, Kimi K3 is highly competitive. For the hardest reasoning chains and multi-domain frontier tasks, Claude Fable 5 and GPT-5.6 Sol still lead.
Kimi K3 has a 1M token context window — the same as Claude Fable 5, Claude Sonnet 5, and Gemini 4. This is large enough to hold approximately 750,000 words, several full-length novels, or a large production codebase with hundreds of files. In practice it means you can feed Kimi K3 an entire repository for review, a full year of company documents, or hundreds of research papers in a single prompt. The large context window is one of Kimi K3's most powerful differentiators for enterprise and developer use cases.
Kimi K3 uses a Mixture-of-Experts (MoE) architecture with 896 expert sub-networks. On each forward pass, the model activates only 16 of those 896 experts per token — roughly 1.8% of the total capacity. This means Kimi K3's effective compute per inference is much lower than a dense 2.8T model would require, making it faster and cheaper to run at inference time. Moonshot introduced two proprietary modifications: Kimi Delta Attention (a hybrid linear attention scheme that reduces memory costs at long context) and Attention Residuals (which change how information flows between layers for better gradient propagation during training).
Yes — all 20 prompts on this page are free to copy and paste directly. Kimi K3 is currently available through the Kimi app (kimi.ai) and Kimi Code. API pricing is $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens. The model's full open weights are also scheduled for public release by July 27, 2026, which will allow self-hosted use at zero per-token cost. Copy any prompt above, paste it into the Kimi app or your API call, and fill in the bracketed placeholders.
Anthropic's most capable model — 20 frontier prompts
Anthropic's new default model with selectable reasoning — 20 prompts
OpenAI's Sol / Terra / Luna tier model — 20 prompts
Microsoft's first in-house reasoning model — 20 prompts
1.6T open-source frontier model — 20 prompts
xAI's latest reasoning model — 20 prompts