The DeepSeek model family at a glance
DeepSeek's models fall into two groups. General models (V4 Pro and V4.1 Flash) handle everyday chat, writing, analysis and coding. Specialist models (R1 for reasoning, Coder for programming) trade breadth for depth in one area. In the chat app you mostly switch between them with the DeepThink toggle; through the API you pick a model by name per request.
V4 Pro
The most capable model for long, nuanced, high-stakes work.
Details → General · defaultV4.1 Flash
Fast and inexpensive; the model behind most chat answers.
Details → Specialist · reasoningR1
Thinks step by step and shows its work; powers DeepThink.
Details → Specialist · codeCoder
Trained on code for completion, generation and repo-level tasks.
Details →DeepSeek V4 Pro - the flagship general model
V4 Pro is DeepSeek's largest and most capable general-purpose model. It produces the most polished long-form writing, handles subtle multi-step analysis, keeps track of very long documents and large codebases, and makes fewer mistakes on hard, open-ended tasks. It is slower and costs more per token than Flash, so it's the model to reach for when quality matters more than speed.
Best for
- Reports, proposals and essays where tone and structure matter
- Deep analysis across long documents or several files
- Complex refactors and architecture questions on large codebases
- Anything you'd otherwise send to a senior specialist
Inputs & access
Text and long documents; file uploads in the chat app; large context for whole-project work. Available in the chat app, the API and Harness Desktop. See our head-to-head: V4 Pro vs V4.1 Flash.
DeepSeek V4.1 Flash - fast, cheap and the default
V4.1 Flash is the workhorse. It answers quickly, costs a fraction of V4 Pro per token, and is the model behind most replies in the chat app. The V4.1-Flash-1017 update (October 2026) significantly improved project-development ability - multi-file coding tasks finish faster and cheaper - and added web search and fetch tools to the API.
Best for
- Everyday chat, drafting, rewriting and summarising
- Customer-facing assistants and high-volume API workloads
- Classification, extraction and other short, repetitive jobs
- Project development in Harness where many quick turns add up
Inputs & access
Text and documents; web search & fetch tool calls through the API. Default model in the chat app; available in the API and Harness. Dedicated page: DeepSeek V4.1 Flash.
DeepSeek R1 - the reasoning model
R1 is a reasoning model: before it answers, it works through the problem in a visible chain of thought - planning, trying an approach, checking the result and backtracking when something doesn't add up. In the chat app it powers the DeepThink toggle; via the API you call it by name and receive the reasoning trace alongside the final answer. R1's weights are open, which made it one of the most widely studied reasoning models released.
Best for
- Maths, logic puzzles and multi-step word problems
- Planning with trade-offs: schedules, budgets, itineraries
- Debugging and tricky algorithmic questions
- Research questions where you want to audit the reasoning
Trade-offs
Slower and more verbose than Flash; overkill for simple rewrites or facts. Inputs: text, plus files via the chat app. Access: DeepThink in chat, the API, or self-hosted from the open weights. Dedicated page: DeepSeek R1.
DeepSeek Coder - built for programming
Coder is trained predominantly on source code across dozens of languages. It excels at code completion (including fill-in-the-middle, where it completes code between a prefix and suffix), generating functions from a docstring, repository-level tasks that need awareness of several files, and translating between languages. It's less conversational than the general models - use it inside tools rather than as a chat companion.
Best for
- IDE autocomplete and inline completions
- Generating boilerplate, tests and documentation
- Refactors and migrations across a codebase
- Code review and explanation pipelines
Inputs & access
Code and text; repository context; FIM prompts through the API. Used by Harness Coding; open weights available for IDE plugins and self-hosting. Dedicated page: DeepSeek Coder.
Side-by-side comparison
Everything above in one table. Strengths are relative within the family.
| Model | Type | Best for / use cases | Supported inputs | Speed | Cost | Where to access |
|---|---|---|---|---|---|---|
| V4 ProFlagship | General | Long-form writing, deep analysis, large codebases, highest-stakes work | Text, long documents, file uploads, large context | Moderate | Highest | Chat · API · Harness |
| V4.1 FlashLatest · 1017 | General | Everyday chat, drafting, summarising, assistants, high-volume API, project development | Text, documents; web search & fetch via API | Fastest | Lowest | Chat (default) · API · Harness |
| R1Reasoning | Specialist | Maths, logic, planning, debugging, auditable research | Text; files via chat; returns reasoning trace | Slow | Moderate | Chat (DeepThink) · API · open weights |
| CoderCode | Specialist | Completion, FIM, generation, repo-level tasks, review pipelines | Code, text, repository context, FIM prompts | Fast | Low | API · Harness Coding · open weights / IDE |
Which DeepSeek model should I use?
Follow the questions. Most people end up on Flash most of the time, and that's the right outcome.
Writing & everyday
Flash. Switch to V4 Pro for the final draft of something important.
Research & analysis
R1 with Search on. Use V4 Pro to write up long findings.
Developers
Coder for completions; Flash for agentic loops; R1 for hard bugs; V4 Pro for architecture.
Where to access each model
Chat app (web, iOS, Android)
Flash by default; DeepThink switches to R1. V4 Pro where offered in the model picker. Free. Download.
How to use →API
All four models by name, OpenAI-compatible, pay per token. Search & fetch tools on Flash. Pricing.
API guide →Harness Desktop
Pick a model per session; Coder drives coding, Flash most document and spreadsheet work.
Harness guide →Self-hosted (open weights)
R1 and Coder weights are published; run them locally or on your own cloud with popular inference servers.
Open weights →Open weights and self-hosting
DeepSeek publishes the weights for most of its models under permissive licences, which is a big part of why the family matters: you can download a model, run it on your own GPUs or a rented cloud instance, fine-tune it on your data, and keep every byte of your prompts in-house.
What's open
- R1 - full weights plus smaller distilled variants that run on a single consumer GPU
- Coder - several sizes, from laptop-friendly to server-class
- Earlier V-series base and chat models
- Check each model's page for the current licence terms and sizes
Why self-host
- Privacy and compliance: data never leaves your infrastructure
- Predictable cost at high volume
- Fine-tuning on your own domain
- No rate limits or "server busy" during peaks
Context windows and supported inputs
A model's context window is how much text it can hold at once - your prompt, any files, the conversation so far and its answer. Bigger windows let you paste whole reports or codebases; smaller ones are cheaper and faster.
| Input | Chat app | API | Harness |
|---|---|---|---|
| Plain text | All models | All models | All models |
| PDF / DOCX / XLSX / TXT | Upload with 📎 | Extract text first, then send | Reads files in your folder directly |
| Images & screenshots | Upload with 📎 | Model-dependent; see API guide | Via plugins |
| Code & repositories | Paste or upload files | Coder / FIM endpoints | Full repo access in Coding mode |
| Live web pages | Search toggle | Search & fetch tools (Flash) | Via plugins |
Pricing at a glance
The chat app is free for all models it offers. The API is prepaid and billed per million tokens, with separate rates for input and output and a discount for cached input. In relative terms:
Lowest
Designed for volume. The sensible default for anything you'll run thousands of times.
Low
Priced for completions in a tight loop; self-hosting is common at scale.
Moderate
Output tokens include the reasoning trace, so long thinking costs more.
Highest
Pay for the best answer when it matters; still inexpensive versus comparable flagships.
Version history
V4.1-Flash-1017
Stronger project development, lower cost, faster responses; web search & fetch added to the API.
V4.1 Flash
The fast, low-cost generation that became the chat default.
V4 Pro
Flagship release focused on long-context quality.
R1
Open-weight reasoning model with visible chain of thought; distilled variants followed.
Coder
Code-specialised family in several sizes with fill-in-the-middle support.
Frequently asked questions about DeepSeek models
What is the best DeepSeek model?
There's no single best. V4 Pro gives the highest quality on long, nuanced tasks; V4.1 Flash is the best everyday choice for speed and cost; R1 is best for reasoning; Coder is best for programming inside tools. See which model should I use.
What's the difference between V4 Pro and V4.1 Flash?
Pro is larger, slower, more expensive and more capable on hard, long tasks. Flash is faster and much cheaper and handles most everyday work just as well. Our V4 Pro vs V4.1 Flash comparison runs the same prompts on both.
Is DeepSeek R1 the same as DeepThink?
DeepThink is the toggle in the chat app; R1 is the model it switches to. Via the API you call R1 directly.
Which DeepSeek model is used in the free chat app?
V4.1 Flash by default, R1 when DeepThink is on, and V4 Pro where the model picker offers it.
Can I run DeepSeek models locally?
Yes for the open-weight models (R1, Coder and earlier V-series). Smaller distilled versions run on a single consumer GPU; full-size models need data-centre hardware. See open weights & self-hosting.
Do DeepSeek models support images?
The chat app accepts image uploads for the models that support them; API support is model-dependent. Check the API guide for the current list.
What is DeepSeek V4.1-Flash-1017?
The October 2026 build of V4.1 Flash, with better project-development capability, lower cost per task, faster responses and web search & fetch in the API. Release notes.
How do DeepSeek models compare with ChatGPT's?
Comparable on everyday tasks, cheaper via API, open-weight, and with visible reasoning in R1; ChatGPT has the larger ecosystem. Full breakdown: DeepSeek vs ChatGPT.