DeepSeek Models Explained: V4 Pro, V4 Flash, R1‑0528, V3.2, Coder & OCR

Last updated: May 3, 2026
Facts last checked: May 3, 2026

DeepSeek Models have evolved from open-weight coding and reasoning models into a broad AI model ecosystem for chat, coding, reasoning, long-context analysis, agent workflows, multimodal research and API-based product development. As of May 2026, the most important models to understand are DeepSeek V4 Pro, DeepSeek V4 Flash, DeepSeek V3/V3.2, DeepSeek R1, and the smaller DeepSeek R1 distilled models. DeepSeek’s official V4 release says the current V4 Preview is available through web, app and API, while the API documentation lists deepseek-v4-pro and deepseek-v4-flash as the current primary model IDs.

Quick answer:
The best DeepSeek Model for most demanding API use is DeepSeek V4 Pro. The best low-cost API option is DeepSeek V4 Flash. For open reasoning research, DeepSeek R1 and its distilled models remain important. For local experimentation, the smaller R1 distilled models are more practical than full-size V4 or V3 models.

DeepSeek Model Picker: Which DeepSeek Model Should You Choose?

Use this quick DeepSeek Model Picker if you already know what you want to build. The best choice depends on whether you need the lowest API cost, the highest current API quality, local reasoning, open-weight research, multimodal understanding, long-context production, or coding-agent workflows.

What you want to do	Choose this DeepSeek model	Why
I want the cheapest API model	DeepSeek V4 Flash	V4 Flash is the economical current API option, designed for fast and cost-efficient production use.
I want the highest API quality	DeepSeek V4 Pro	V4 Pro is the stronger current API model for complex reasoning, coding, agent workflows and high-value tasks.
I want local reasoning	DeepSeek R1 Distill 14B or 32B	The R1 distilled models are more practical for local reasoning experiments than full-size R1, V3 or V4 models.
I want open-weight reasoning research	DeepSeek R1 / R1-0528 / V3.2	R1 and R1-0528 are useful for reasoning research, while V3.2 is useful for studying DeepSeek’s open-weight MoE and agentic model evolution.
I want multimodal capabilities	DeepSeek VL2 / Janus / DeepSeek OCR	These specialized models are better suited for vision-language, image understanding, generation or OCR tasks than text-only V4 models.
I want long-context production	DeepSeek V4 Pro or DeepSeek V4 Flash	Both current V4 API models support a 1M-token context window, making them suitable for long documents, codebases and knowledge workflows.
I want a coding agent	DeepSeek V4 Pro	V4 Pro is the safer choice for complex agentic coding, multi-step software tasks and tool-heavy developer workflows.
I want a general chatbot	DeepSeek V4 Flash	V4 Flash is a strong default for everyday chat, support bots and high-volume conversational workloads where cost matters.
I want research into older DeepSeek architecture	DeepSeek V3 / V3.2	V3 and V3.2 are useful for understanding DeepSeekMoE, Multi-head Latent Attention, Sparse Attention and the evolution toward later reasoning and agent models.

Simple rule: start with V4 Flash when cost matters, use V4 Pro when quality matters, choose R1 Distill for local reasoning, use R1 or R1-0528 for open reasoning research, and choose VL2, Janus or OCR for multimodal tasks.

What Are DeepSeek Models?

DeepSeek Models are AI models developed by DeepSeek, a Chinese AI company focused on building advanced artificial intelligence systems. The DeepSeek ecosystem includes large language models for chat and reasoning, code models, distilled reasoning models, multimodal models, OCR-focused models and formal theorem-proving models. DeepSeek’s official website links to research families including R1, V3, Coder V2, VL, V2, Coder, Math and LLM.

In practical terms, a DeepSeek Model can be used for tasks such as writing, summarization, coding, software-agent workflows, mathematical reasoning, tool calling, document analysis and research. The newer V4 models are especially important because they introduce a 1-million-token context window across official DeepSeek services and are available through both OpenAI-compatible and Anthropic-compatible API formats.

DeepSeek’s model lineup is not a single model. It is a family of models optimized for different trade-offs:

• V4 Pro: strongest current general-purpose DeepSeek model for complex reasoning, coding and agents.
• V4 Flash: faster and cheaper current API model for high-volume use.
• V3/V3.2: important open-weight predecessor models that introduced major efficiency and agentic improvements.
• R1: reasoning-focused family trained around reinforcement learning and chain-of-thought-style problem solving.
• R1 distilled models: smaller dense models based on Qwen and Llama, useful for local reasoning experiments.
• Coder, VL, Janus, OCR and Prover: specialized model families for code, vision-language, image generation, OCR and formal proof work.

Quick DeepSeek Model Comparison

Model	Type	Best for	Strengths	Limitations	Access options
DeepSeek V4 Pro	MoE language model	Complex reasoning, agentic coding, long-context work, high-value API tasks	1.6T total parameters, 49B active parameters, 1M context, thinking and non-thinking modes	Higher API cost than Flash; very large for self-hosting	Web, app, API, Hugging Face open weights
DeepSeek V4 Flash	Smaller MoE language model	Low-cost API usage, general chat, volume workloads, fast responses	284B total parameters, 13B active parameters, 1M context, lower pricing	Weaker than Pro for the hardest knowledge and agentic tasks	Web, app, API, Hugging Face open weights
DeepSeek V3.2	MoE language model	Research, historical comparison, agentic reasoning, open-weight experimentation	DeepSeek Sparse Attention, thinking with tool use, MIT license	No longer the primary current API model after V4	Hugging Face, GitHub, research use
DeepSeek V3	MoE language model	Baseline open-weight LLM research, general chat and coding comparisons	671B total parameters, 37B active per token, MLA and DeepSeekMoE	Superseded by V3.2 and V4 for most current use cases	GitHub, Hugging Face
DeepSeek R1	Reasoning-focused MoE model	Math, logic, code reasoning, research into RL-based reasoning	671B total parameters, 37B active parameters, 128K context	Older than V4; full model is large for local deployment	GitHub, Hugging Face, research
DeepSeek R1-Zero	RL-first reasoning model	Research into reinforcement learning without initial SFT	Demonstrates self-verification, reflection and long reasoning behavior	Less aligned and less polished than R1 for general use	GitHub, Hugging Face
DeepSeek R1 Distill models	Smaller dense reasoning models	Local experimentation, smaller deployments, education, research	1.5B to 70B checkpoints based on Qwen and Llama	Not equivalent to full R1 or V4; performance depends on size	Hugging Face
DeepSeek Coder V2	Code-specialized MoE model	Code generation, code completion, software engineering tasks	16B and 236B variants, 128K context, expanded programming language support	Older specialized family; V4 may be preferable for modern agentic coding	GitHub, Hugging Face, platform references
DeepSeek VL2 / Janus / OCR / Prover	Specialized multimodal, OCR and proof models	Vision-language, image generation, OCR, formal theorem proving	Dedicated capabilities beyond text-only LLMs	Not replacements for general V4 API chat	GitHub, Hugging Face

The V4 parameter counts, 1M context length and API availability come from DeepSeek’s official V4 release and model card. V3, V3.2, R1, Coder V2 and the specialized model families are documented in official DeepSeek GitHub or Hugging Face pages.

Latest DeepSeek Models: V4 Pro and V4 Flash

The latest major DeepSeek Models are DeepSeek V4 Pro and DeepSeek V4 Flash, released as part of DeepSeek V4 Preview on April 24, 2026. DeepSeek describes the V4 Preview as open-sourced and available through chat, app and API. It also says V4 introduces a default 1M context length across official DeepSeek services.

DeepSeek V4 Pro

DeepSeek V4 Pro is the stronger current DeepSeek API model. According to DeepSeek, it has 1.6 trillion total parameters and 49 billion active parameters. It is designed for complex reasoning, knowledge-heavy tasks, agentic coding and high-value workflows where accuracy and depth matter more than raw cost.

DeepSeek V4 Flash

DeepSeek V4 Flash is the more economical current DeepSeek API model. According to DeepSeek, it has 284 billion total parameters and 13 billion active parameters. It is positioned as faster, more efficient and more cost-effective than V4 Pro, while still supporting the same 1M context length and API feature set.

DeepSeek V4 Pro vs V4 Flash

Feature	DeepSeek V4 Pro	DeepSeek V4 Flash
API model ID	deepseek-v4-pro	deepseek-v4-flash
Total parameters	1.6T	284B
Active parameters	49B	13B
Context length	1M tokens	1M tokens
Max output	Up to 384K tokens	Up to 384K tokens
Thinking mode	Supported	Supported
Non-thinking mode	Supported	Supported
JSON output	Supported	Supported
Tool calls	Supported	Supported
Chat Prefix Completion	Supported	Supported
FIM Completion	Non-thinking mode only	Non-thinking mode only
Best fit	High-value reasoning, coding, agent tasks	Low-cost production, general chat, high-volume workloads

The API documentation lists both V4 models with 1M context, maximum output of 384K tokens, JSON output, tool calls, Chat Prefix Completion and FIM Completion in non-thinking mode only.

DeepSeek V3 and V3.2: Why They Still Matter

DeepSeek V3 is still important because it established much of the modern DeepSeek architecture. The official DeepSeek V3 GitHub page describes it as a Mixture-of-Experts model with 671B total parameters and 37B activated for each token. It also says V3 uses Multi-head Latent Attention and DeepSeekMoE, and that it was pretrained on 14.8 trillion tokens.

DeepSeek V3.2 matters because it introduced additional efficiency and agentic improvements before V4. The Hugging Face model card describes V3.2 as a model focused on efficient reasoning and agentic AI, with DeepSeek Sparse Attention, a scalable reinforcement learning framework and a large-scale agentic task synthesis pipeline. The same page lists the V3.2 model weights under the MIT license.

DeepSeek’s API history also makes V3.2 relevant. On December 1, 2025, DeepSeek said deepseek-chat and deepseek-reasoner had been upgraded to V3.2, with deepseek-chat corresponding to non-thinking mode and deepseek-reasoner corresponding to thinking mode. On April 24, 2026, the change log said those legacy names now point to V4 Flash modes and will be discontinued on July 24, 2026.

For current API users, this means V3.2 is mostly a historical and open-weight reference. For new applications, use deepseek-v4-pro or deepseek-v4-flash directly instead of relying on legacy aliases.

DeepSeek R1: The Reasoning-Focused Model Family

DeepSeek R1 is the reasoning-focused DeepSeek Model family. DeepSeek released R1 in January 2025 and described it as fully open-source, with code and models under the MIT license. The release also introduced six open-source distilled models.

The R1 GitHub repository explains that DeepSeek R1-Zero was trained by applying reinforcement learning directly to the base model without an initial supervised fine-tuning stage. DeepSeek says this produced behaviors such as self-verification, reflection and long chain-of-thought-style reasoning. The same repository describes DeepSeek R1 as using a pipeline with two RL stages and two supervised fine-tuning stages.

The full R1 and R1-Zero models are both listed as 671B total parameters, 37B active parameters and 128K context length, and both are trained based on DeepSeek V3-Base.

Use R1-style models when the main job is reasoning rather than ordinary chat. Good examples include math problems, logic puzzles, code reasoning, algorithm design, structured planning and research into reinforcement learning for language models. For most new production API workflows, however, V4 Pro and V4 Flash are now the more current choices.

Distilled DeepSeek Models

Distillation means training a smaller model to imitate useful behaviors from a larger model. In the DeepSeek R1 family, DeepSeek generated reasoning data from R1 and fine-tuned smaller dense models based on Qwen and Llama. The official R1 repository says the distilled checkpoints include 1.5B, 7B, 8B, 14B, 32B and 70B models.

Distilled model	Base model	Size	Best use case	Local deployment difficulty
DeepSeek-R1-Distill-Qwen-1.5B	Qwen2.5-Math-1.5B	1.5B	Education, lightweight reasoning experiments	Low
DeepSeek-R1-Distill-Qwen-7B	Qwen2.5-Math-7B	7B	Small local reasoning, notebooks, prototypes	Low to medium
DeepSeek-R1-Distill-Llama-8B	Llama-3.1-8B	8B	Local reasoning with Llama ecosystem compatibility	Low to medium
DeepSeek-R1-Distill-Qwen-14B	Qwen2.5-14B	14B	Stronger local reasoning experiments	Medium
DeepSeek-R1-Distill-Qwen-32B	Qwen2.5-32B	32B	Advanced local reasoning, research labs	Medium to high
DeepSeek-R1-Distill-Llama-70B	Llama-3.3-70B-Instruct	70B	High-end local or hosted reasoning workloads	High

The distilled models are not simply “small DeepSeek R1.” They are smaller base models fine-tuned on R1-generated reasoning data. That makes them useful when you need reasoning behavior but cannot run a full 671B-parameter MoE model.

Which DeepSeek Model Should You Use?

Use case	Recommended DeepSeek model	Why
General chatbot	DeepSeek V4 Flash	Lower API cost, fast responses, strong general capability
Premium chatbot	DeepSeek V4 Pro	Better for complex user requests, deeper reasoning and higher-value answers
Long-context document analysis	DeepSeek V4 Pro or V4 Flash	Both support 1M context through the current API
Coding assistant	DeepSeek V4 Pro	Strongest current choice for agentic coding and complex software tasks
High-volume coding support	DeepSeek V4 Flash	More economical for repeated coding help and simpler agent tasks
Agentic workflows	DeepSeek V4 Pro	Best current DeepSeek choice for tool-heavy, multi-step workflows
Math and reasoning	DeepSeek V4 Pro with thinking enabled, or DeepSeek R1 for research	V4 is current for API; R1 remains important for open reasoning research
Low-cost API usage	DeepSeek V4 Flash	Lowest listed current V4 API price
Local experimentation	R1 distilled models	More practical sizes than full V4, V3 or R1
Enterprise deployment	V4 Pro via API or open weights depending on governance needs	Strong current model; deployment choice depends on privacy, cost and infrastructure
Research	V3.2, V4, R1, R1-Zero and distilled R1	Open weights and model cards support reproducible comparison
Multimodal or OCR tasks	DeepSeek VL2, Janus or DeepSeek OCR	Dedicated model families for vision-language, image generation and OCR

For API-first products, start with V4 Flash for cost-sensitive workloads and upgrade difficult tasks to V4 Pro. For research and local model work, evaluate R1 distilled models first before attempting full-size MoE deployment.

DeepSeek API Model IDs, Pricing and Context Length

Last checked: May 3, 2026

Model ID	Model version	Context length	Max output	Input price, cache hit	Input price, cache miss	Output price	Key features
deepseek-v4-flash	DeepSeek V4 Flash	1M	384K	$0.0028 / 1M tokens	$0.14 / 1M tokens	$0.28 / 1M tokens	Thinking/non-thinking, JSON output, tool calls, Chat Prefix Completion, FIM in non-thinking mode
deepseek-v4-pro	DeepSeek V4 Pro	1M	384K	$0.003625 / 1M tokens during discount	$0.435 / 1M tokens during discount	$0.87 / 1M tokens during discount	Thinking/non-thinking, JSON output, tool calls, Chat Prefix Completion, FIM in non-thinking mode
deepseek-chat	Legacy alias	Not recommended for new apps	Not recommended	Routes to V4 Flash non-thinking mode	Routes to V4 Flash non-thinking mode	Routes to V4 Flash non-thinking mode	Deprecated alias; avoid for new builds
deepseek-reasoner	Legacy alias	Not recommended for new apps	Not recommended	Routes to V4 Flash thinking mode	Routes to V4 Flash thinking mode	Routes to V4 Flash thinking mode	Deprecated alias; avoid for new builds

DeepSeek’s pricing page says prices are listed per 1M tokens, that product prices may vary, and that users should regularly check the pricing page for current information. It also says deepseek-chat and deepseek-reasoner correspond to V4 Flash non-thinking and thinking modes for compatibility, while the change log says those legacy aliases will be discontinued on July 24, 2026.

The current V4 Pro prices shown above reflect DeepSeek’s listed 75% discount, which the pricing page says is extended until May 31, 2026, 15:59 UTC. The same page says cache-hit input prices were reduced to one-tenth of launch price effective April 26, 2026, 12:15 UTC.

Production warning: Do not hard-code pricing assumptions into business models. Always re-check DeepSeek’s official pricing page before estimating customer margins, setting token budgets or signing enterprise commitments.

How to Access DeepSeek Models

1. DeepSeek web and app

DeepSeek’s official website says V4 Preview is available on web, app and API. For non-developer users, the web or app experience is the fastest way to test the latest DeepSeek Models without building an integration.

2. DeepSeek API

The DeepSeek API supports OpenAI-compatible and Anthropic-compatible formats. The quick-start documentation lists deepseek-v4-flash, deepseek-v4-pro, deepseek-chat and deepseek-reasoner, while warning that the two legacy names will be deprecated on July 24, 2026.

Example OpenAI-compatible call:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a helpful technical assistant."},
        {"role": "user", "content": "Explain DeepSeek V4 in simple terms."}
    ],
    extra_body={"thinking": {"type": "enabled"}},
    reasoning_effort="high",
    stream=False
)

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

3. Hugging Face model cards

DeepSeek hosts many model weights and model cards on Hugging Face. The DeepSeek Hugging Face organization lists the DeepSeek V4 collection, including V4 Flash Base, V4 Flash, V4 Pro Base and V4 Pro.

4. GitHub repositories

DeepSeek also publishes major research repositories on GitHub, including V3, R1, Coder V2, VL, Janus, OCR and Prover. These repositories are useful for model details, papers, inference instructions, release notes and research context.

5. Coding and agent tools

DeepSeek documents integrations with AI coding and agent tools. The API docs include guidance for Claude Code, OpenCode and OpenClaw, and a separate page describes a VS Code extension that adds DeepSeek V4 Pro and V4 Flash to GitHub Copilot Chat’s model picker.

6. Self-hosting and local deployment

Self-hosting depends heavily on model size and infrastructure. V4 Pro and V4 Flash are open-weight models, but their total parameter counts are extremely large. DeepSeek’s V4 Hugging Face card provides local inference guidance and says Think Max mode should use a context window of at least 384K tokens. This makes full V4 self-hosting a high-infrastructure task rather than a typical consumer PC setup.

For local experimentation, start with R1 distilled models or smaller specialized models before attempting full V3, R1 or V4 deployments.

Technical Concepts Behind DeepSeek Models

Mixture-of-Experts

Many major DeepSeek Models use a Mixture-of-Experts architecture. Instead of activating every parameter for every token, an MoE model routes each token through a subset of expert parameters. This is why V4 Pro can have 1.6T total parameters but only 49B active parameters per token, and why V4 Flash can have 284B total parameters but 13B active parameters.

Active parameters

Active parameters are the parameters used for a given token during inference. They matter because they help explain the compute trade-off of MoE models. A model can be very large overall while using a smaller active subset per token.

Multi-head Latent Attention and DeepSeekMoE

DeepSeek V3 adopted Multi-head Latent Attention and DeepSeekMoE for efficient inference and cost-effective training. These concepts are important because V3 became the base architecture for later reasoning work, including R1.

DeepSeek Sparse Attention

DeepSeek V3.2 introduced DeepSeek Sparse Attention, described by DeepSeek as an efficient attention mechanism for reducing computational complexity while preserving model performance in long-context scenarios. V4 later incorporated additional attention innovations for million-token context efficiency.

Thinking mode

Thinking mode controls whether the model uses an explicit reasoning-oriented mode. In the current API, the thinking object can be enabled or disabled, and the reasoning_effort field supports high and max. The default is thinking enabled.

Tool calls

Tool calls allow the model to call external functions or tools, such as retrieval, calculators, internal business systems or coding agents. Both current V4 API models support tool calls according to the pricing/model-details page.

Context caching

Context caching reduces cost when repeated input context is reused. DeepSeek’s pricing table separates cache-hit and cache-miss input token prices, which is important for long-context products such as document assistants, codebase assistants and knowledge-base agents.

FIM Completion

FIM means fill-in-the-middle, a common capability for code completion where the model fills missing content between a prefix and suffix. DeepSeek’s current V4 pricing table says FIM Completion is supported in non-thinking mode only.

Distillation

Distillation transfers useful behavior from a larger model into a smaller model. DeepSeek used R1-generated reasoning data to fine-tune smaller Qwen- and Llama-based models, creating the R1 distilled family.

DeepSeek Models vs Other AI Models

DeepSeek Models are best compared by category rather than by unsupported “winner takes all” claims.

Category	How DeepSeek compares
Cost	V4 Flash is positioned as the economical current API model, while V4 Pro is the stronger but more expensive option.
Open-weight availability	DeepSeek publishes many model weights, including V4, V3.2, V3, R1 and distilled R1 models.
Reasoning	R1 established DeepSeek’s reasoning reputation, while V4 adds current thinking modes through the API.
Coding	Coder V2 was the specialized coding family; V4 Pro is now the stronger current option for agentic coding workflows.
Context length	Current V4 API models support a 1M-token context window.
Deployment flexibility	Developers can use the API, Hugging Face weights, GitHub repositories or coding-agent integrations.
Governance	API use is simpler; open-weight use gives more deployment control but requires infrastructure, safety review and operational expertise.

The strongest DeepSeek advantage is not one single benchmark. It is the combination of open-weight releases, low listed API prices, long-context support and a model family that spans general chat, reasoning, coding and specialized research. The main trade-off is that the largest models are difficult to self-host, and model names, API aliases and pricing have changed over time.

Limitations and Risks

1. The lineup changes quickly

DeepSeek has changed API mappings over time. For example, deepseek-chat and deepseek-reasoner moved from V3.2 mappings to V4 Flash compatibility mappings, and DeepSeek says those names will be discontinued on July 24, 2026.

2. Pricing can change

DeepSeek explicitly says product prices may vary and recommends checking the pricing page regularly. This matters for startups, SaaS products and agents that process large token volumes.

3. V4 is text-only

DeepSeek’s GitHub Copilot integration page says DeepSeek V4 is text-only and that the extension handles images by routing image descriptions through another installed model before sending text to DeepSeek. For native multimodal work, use DeepSeek VL2, Janus or OCR-family models instead.

4. Full-size self-hosting is not simple

V4 Pro, V4 Flash, V3 and R1 are very large models. Even when weights are available, production self-hosting requires careful planning around GPUs, quantization, serving software, memory, context length, throughput and monitoring. This is an infrastructure project, not just a model download.

5. Benchmarks need context

DeepSeek publishes many benchmark claims, but production quality depends on your data, prompts, latency requirements, safety needs and evaluation method. Treat vendor benchmarks as useful signals, not as a substitute for your own testing.

6. Hallucination and safety risks remain

Like other LLMs, DeepSeek Models can produce incorrect, incomplete or misleading outputs. Use retrieval, validation, human review and tool-level safeguards for legal, financial, medical, security or high-impact decisions.

7. Privacy and jurisdiction matter

Before sending private data to any API, review your data governance requirements, regional regulations, retention expectations and vendor terms. For sensitive workloads, compare API usage with self-hosted or private deployment options.

Final Recommendation

For most new users, the best starting point is simple:

• Best overall DeepSeek Model: DeepSeek V4 Pro.
• Best low-cost DeepSeek Model: DeepSeek V4 Flash.
• Best model for reasoning research: DeepSeek R1 or R1-Zero.
• Best local experimentation option: DeepSeek R1 distilled models.
• Best developer/API option: V4 Pro for high-value tasks, V4 Flash for scalable production workloads.
• Best multimodal/OCR option: DeepSeek VL2, Janus or DeepSeek OCR, depending on the task.

If you are building a production app today, use the explicit current model IDs: deepseek-v4-pro and deepseek-v4-flash. Avoid relying on deepseek-chat or deepseek-reasoner for new projects because DeepSeek has already marked those aliases for deprecation.

FAQ