DeepSeek Models Explained: V4, R1, V3 and Distilled Models

by David Porter

Tuesday, 11 August 2026 at 02:00

thumbnail_deepseek-models-explained-v4-r

DeepSeek has released several model families in a short period: V3 established its efficiency story, R1 popularized its reasoning approach, distilled R1 variants made local experimentation easier, and V4 moved the hosted platform toward one-million-token context and agentic work.

As of August 7, 2026, the current official API model IDs are deepseek-v4-flash and deepseek-v4-pro. The older API aliases deepseek-chat and deepseek-reasoner were retired on July 24, 2026. R1 and V3 still matter as downloadable checkpoints and in existing deployments, but they should not be presented as the current hosted default.

This guide owns the model map. For the complete company and product overview, start with DeepSeek AI: Complete Guide. For exact rates, use our DeepSeek pricing guide.

DeepSeek models at a glance

Model or family	Current role	Size and context	Access	Best fit
DeepSeek V4 Pro	Current high-capability hosted model	1.6T total parameters, 49B active; 1M context	DeepSeek API; V4 open weights/repositories	Difficult reasoning, knowledge work, long documents and complex agents
DeepSeek V4 Flash 0731	Current efficiency and agent model	284B total, 13B active in the core architecture; 1M context	API, Responses API and open weights	High-volume coding, agents, transformations and cost-sensitive workloads
DeepSeek R1	Reasoning checkpoint	671B total, 37B active; 128K in official repository	Downloadable weights and third-party hosts	Research, reasoning studies and large self-hosted deployments
R1 distilled models	Smaller reasoning-focused variants	1.5B, 7B, 8B, 14B, 32B and 70B families	Local tools and third-party repositories	Desktop or workstation experimentation
DeepSeek V3 / V3.2	Earlier general model family	V3: 671B total, 37B active; 128K	Downloadable weights, historical deployments	Research, legacy systems and architecture study
DeepSeek Coder families	Earlier coding-specialist releases	Varies by release	Downloadable repositories and older integrations	Legacy code tasks and research; superseded in the hosted product by newer general/agent models

The numbers require interpretation. In a mixture-of-experts model, total parameters describe the full collection of experts stored in the checkpoint, while active parameters describe the smaller subset used for a token. Fewer active parameters can reduce inference computation, but the complete model still requires substantial storage and distributed serving infrastructure.

DeepSeek V4: the current platform generation

DeepSeek introduced V4 in April 2026 with two variants: Pro and Flash. Both were designed around long context, thinking and non-thinking operation, and stronger agent capabilities.

The official V4 release page lists:

1.6 trillion total parameters and 49 billion active parameters for V4 Pro;
284 billion total parameters and 13 billion active parameters for V4 Flash;
a one-million-token context window across official V4 services;
OpenAI Chat Completions and Anthropic-compatible API access;
thinking and non-thinking modes;
and open model weights.

The current pricing and model table additionally lists a maximum output of 384,000 tokens, JSON output and tool calls for both hosted models.

V4 Pro

V4 Pro is positioned as the more capable option for difficult reasoning, broad knowledge and demanding agent workflows. It has the larger expert system and the higher API price.

Choose Pro when:

an error is more expensive than the additional token cost;
the task requires difficult planning or multi-step reasoning;
long documents contain subtle dependencies;
an agent needs to make consequential tool choices;
or Flash repeatedly fails a representative evaluation.

Do not choose Pro automatically for every request. A short classification, metadata extraction or simple rewrite may gain little from the larger model while paying more for output and latency.

V4 Flash

V4 Flash is the lower-cost and faster V4 variant. “Flash” means efficient relative to V4 Pro; it does not mean small enough for an ordinary laptop. DeepSeek’s own model card and serving examples place the full checkpoint firmly in multi-GPU territory.

The current API points deepseek-v4-flash to DeepSeek-V4-Flash-0731. The July 31 update kept the same core architecture and size as the preview but applied new post-training, with a particular emphasis on coding agents and tool use.

Flash is the best first test for:

high-volume summarization and transformation;
code explanation and routine generation;
repository agents with strong verification;
structured extraction;
long-context filtering;
and workloads whose economics matter at large scale.

At the August 7 snapshot, Flash is also the only DeepSeek model supported by the Responses API. V4 Pro supports Chat Completions and the Anthropic-compatible interface, but not yet Responses.

The July 2026 V4 Flash update: API versus app

DeepSeek’s change log makes an important distinction. The July 31, 2026 update upgraded the V4 Flash API. The company explicitly said the V4 Pro API and the app/web models were unchanged by that release.

That means “DeepSeek uses V4 Flash 0731” cannot be applied indiscriminately to every interface. The consumer product may expose modes or versions on a different schedule. Developers should identify the exact API model, while consumer reviewers should report the interface and date tested.

A reliable model record includes:

provider;
endpoint;
model ID;
repository or checkpoint name where relevant;
date tested;
reasoning effort;
system instructions;
tool configuration;
and evaluation set.

Without that information, model comparisons are difficult to reproduce.

What does one-million-token context mean?

A context window is the maximum working space available to a request. It includes system instructions, user messages, previous conversation, retrieved documents, tool results and generated content counted under the provider’s rules.

A one-million-token window can hold an enormous amount of text. It can support:

analysis of a large code repository;
comparison of many contracts or policy documents;
long research archives;
multi-stage agents retaining extensive state;
and extraction across large document collections.

It does not mean the model will reason perfectly across every token. Four constraints remain.

Retrieval quality

A model may overlook a relevant passage in a huge input. Retrieval can still improve quality by selecting the most relevant evidence before generation.

Cost

A large uncached prompt can be inexpensive by frontier-model standards and still become material at high volume. The DeepSeek pricing guide shows worked examples.

Latency

Processing more tokens takes time. Long context can make an interactive application feel slow even when the model supports it technically.

Memory and serving

Self-hosted long context increases key-value cache and memory requirements. A checkpoint that fits at short context may not sustain the desired concurrency at one million tokens.

Thinking and non-thinking modes

V4 Pro and Flash support thinking and non-thinking behavior through the hosted API.

Non-thinking mode is suitable for:

direct transformations;
simple extraction;
routine customer-service drafts;
formatting;
and tasks where speed matters more than deliberation.

Thinking mode is useful for:

difficult coding;
multi-step planning;
mathematics;
ambiguous analysis;
and agent workflows that require a sequence of decisions.

The API exposes reasoning-effort controls, but the mapping differs by model and can change. At the cut-off date, V4 Flash supported low, high and max behavior more directly, while V4 Pro temporarily mapped some requested levels upward. Production systems should test output quality, token use and latency instead of assuming identical effort semantics.

Thinking mode should not be treated as a guarantee of truth. A model can reason coherently from a false premise or invent a source. Applications should rely on final answers, observable tool calls and verifiable evidence rather than exposing hidden reasoning as a trust mechanism.

DeepSeek R1: why it still matters

DeepSeek R1 was released in January 2025 and became one of the company’s most influential models. It drew attention to reinforcement-learning-based reasoning and made a powerful reasoning checkpoint available for download.

The official DeepSeek R1 repository lists the full R1 and R1-Zero models at:

671 billion total parameters;
37 billion active parameters;
and a 128,000-token context window.

R1 is no longer the current broad hosted API name. The legacy deepseek-reasoner alias was retired after being temporarily remapped through newer generations. R1 remains relevant for:

academic and independent research;
organizations that already deployed the checkpoint;
studies of reasoning behavior;
third-party inference services;
and distilled models for local use.

R1 versus R1-Zero

R1-Zero was trained to explore reasoning behavior with reinforcement learning and less conventional supervised preparation. It demonstrated striking behavior but also produced issues such as readability and language mixing. DeepSeek R1 added training stages intended to improve usefulness and presentation.

For ordinary use, R1 is the practical reference. R1-Zero is primarily a research artifact.

R1 distilled models

The full 671B R1 model is far too large for normal desktop use. DeepSeek therefore released distilled variants based on smaller Qwen and Llama architectures.

The official repository lists:

DeepSeek-R1-Distill-Qwen-1.5B;
DeepSeek-R1-Distill-Qwen-7B;
DeepSeek-R1-Distill-Llama-8B;
DeepSeek-R1-Distill-Qwen-14B;
DeepSeek-R1-Distill-Qwen-32B;
DeepSeek-R1-Distill-Llama-70B.

A distilled model is not the full R1 compressed into a smaller file. It is a different base model trained or fine-tuned using outputs and reasoning examples from the larger teacher. It can inherit useful behavior without matching the teacher’s full capability.

Distills are attractive because they create realistic local tiers:

Distilled size	Practical interpretation
1.5B	Very light experimentation; fast but limited
7B–8B	Entry-level local reasoning and coding on modest hardware
14B	Better quality with a meaningful memory increase
32B	Workstation-class local model for stronger output
70B	High-end workstation or multi-GPU/server use

Actual memory depends on quantization, runtime, context length, concurrency and cache. File size is not the full memory requirement.

Our local guide explains how to run DeepSeek with Ollama, LM Studio and vLLM.

DeepSeek V3 and V3.2

DeepSeek V3 established the architecture that underpinned R1 and much of the company’s global reputation. Its official repository describes a 671B-parameter mixture-of-experts model with 37B active parameters and a 128K context window.

V3 was notable for:

mixture-of-experts routing;
Multi-Head Latent Attention;
aggressive training-efficiency claims;
strong coding, math and general-language results;
and an open-weight release.

DeepSeek later updated the family through releases including V3.1 and V3.2. V3.2 emphasized reasoning and agent capabilities before V4 became the current generation.

V3 remains important when:

reproducing older research;
supporting an existing deployment;
comparing architectural generations;
or using a third-party provider that has not migrated.

A new API integration should not begin with the old deepseek-chat alias. Use the current V4 IDs and the DeepSeek API guide.

What happened to DeepSeek Coder?

DeepSeek released dedicated coding-model families before newer general models integrated strong code and agent capabilities. DeepSeek Coder and Coder V2 helped establish the company’s reputation among developers.

The hosted product later merged coding and general capabilities into broader model generations. Older Coder repositories may still be useful for:

research on specialist models;
constrained local deployment;
compatibility with an existing stack;
and tasks where a known older checkpoint has already been evaluated.

For a new production system, V4 Flash or Pro is normally the more relevant hosted choice. Do not assume a model with “Coder” in its name will outperform a newer general model on repository agents, tool use or your language stack.

API model IDs versus repository names

DeepSeek uses stable API identifiers and more specific checkpoint names.

Type	Example	Meaning
API model ID	deepseek-v4-flash	Provider-managed alias used in requests
Current mapped version	DeepSeek-V4-Flash-0731	Specific model behind the alias at the cut-off date
Repository name	deepseek-ai/DeepSeek-V4-Flash-0731	Downloadable checkpoint and model card
Legacy API alias	deepseek-chat	Retired July 24, 2026
Historical checkpoint	DeepSeek-R1	Fixed downloadable release, not the current hosted default

A provider-managed alias is convenient because applications do not need to change names for every update. It also creates change risk: output behavior may move when the provider remaps the alias.

Production applications should maintain regression tests and record the date of each provider update. Where the API offers a pinned version, use it when reproducibility matters. When only a rolling alias exists, create an internal approval process for model changes.

Are DeepSeek models open source?

DeepSeek frequently describes its releases as open source. For model articles, open-weight is the more precise default.

Open weights mean developers can download the learned parameters and run the model under the stated license. That provides meaningful benefits:

independent evaluation;
private or sovereign deployment;
fine-tuning where permitted;
model portability;
and reduced dependence on one API.

It does not necessarily provide:

the complete training dataset;
every data-cleaning decision;
exact reproducibility of the training run;
the full internal evaluation suite;
or the hosted service’s post-training and safety stack.

The current V4 Flash repository displays an MIT license. R1’s repository also uses MIT for the main release, while distilled models inherit licensing considerations from their base models. Always review the exact repository and downstream license before commercial use.

How to choose a DeepSeek model

Use a staged selection process rather than choosing the largest model first.

Step 1: choose hosted or self-hosted

Use the official API when speed of implementation and low initial operational burden matter. Use self-hosting when data control, offline operation, customization or model portability justifies infrastructure work.

Step 2: establish a quality threshold

Create representative tasks and a scoring rubric. Include failure cases, not only ideal prompts.

Step 3: test Flash first for hosted workloads

Flash’s current price and agent performance make it the natural baseline. Measure whether it meets the threshold.

Step 4: escalate failures to Pro

Send only harder requests to Pro when the additional quality is measurable. A router can keep routine work on Flash and reserve Pro for complex cases.

Step 5: choose a local tier by memory and quality

For a desktop, begin with an 8B or 14B distilled R1 model. Move to 32B only if the workstation can support it and the quality gain matters. Treat 70B and full models as server projects.

Step 6: version everything

Record the model, runtime, quantization, context, prompt and test result. “DeepSeek was good” is not an actionable evaluation.

Model recommendations by workload

Workload	Starting choice	Why
Routine API extraction	V4 Flash, non-thinking	Lowest cost and straightforward output
Coding agent	V4 Flash 0731 with tools and verification	Updated specifically for agentic work
Difficult reasoning	V4 Pro thinking mode	Higher-capability hosted route
Very long document review	Flash baseline, Pro escalation	Both support 1M context; route by difficulty
Local laptop experiment	R1 distill 7B or 8B, quantized	Realistic memory footprint
Local workstation	R1 distill 14B or 32B	Better quality with manageable hardware for advanced users
Research on reasoning	Full R1 or R1-Zero where infrastructure permits	Direct access to original checkpoints
Reproducing 2024–2025 work	Matching V3/R1 checkpoint	Keeps historical comparisons valid
New general production API	V4 Flash or Pro	Current official model IDs

How to evaluate models without being misled by benchmarks

Benchmark tables are useful evidence, but they are not universal rankings.

Check:

whether results are vendor-reported or independently reproduced;
whether the model used tools, search or an agent harness;
reasoning effort and output-token budget;
number of attempts or sampling strategy;
whether the test set may be contaminated;
latency and cost;
and whether your real task resembles the benchmark.

DeepSeek’s V4 Flash model card, for example, states that coding-agent benchmarks used a particular harness, maximum reasoning effort and specified sampling settings. Those details matter. A result from that configuration cannot be assumed for a simple chat prompt with no tools.

The best model is the least expensive system that reliably completes the defined task under your actual controls.

Frequently asked questions

What is the latest DeepSeek model?

The current hosted API family is DeepSeek V4. deepseek-v4-flash maps to V4 Flash 0731, while deepseek-v4-pro provides the larger Pro route at the August 7, 2026 snapshot.

Is DeepSeek R1 still available?

Yes as a downloadable checkpoint and through some third-party services. It is not the current default official API model name.

What replaced deepseek-chat?

DeepSeek retired deepseek-chat and deepseek-reasoner on July 24, 2026. New integrations should use deepseek-v4-flash or deepseek-v4-pro.

Which DeepSeek model is best for coding?

V4 Flash 0731 is the strongest default for cost-sensitive coding agents because the update emphasized agentic coding and tool use. Pro may perform better on particular difficult tasks. Test both on your repositories and require tests and review.

Which DeepSeek model can run on a laptop?

Small quantized R1 distilled models—typically 1.5B, 7B or 8B—are the realistic starting point. Full R1, V3 and V4 are not normal laptop models.

What is the difference between V4 Pro and V4 Flash?

Pro is the larger, more expensive model positioned for difficult reasoning and knowledge work. Flash is smaller relative to Pro, much cheaper through the API and currently has the updated agentic checkpoint plus Responses API support.

Does V4 have a one-million-token context window?

DeepSeek’s current official hosted model table lists one million tokens for both V4 Pro and Flash. Effective quality across very large contexts still depends on retrieval, prompt structure and the task.

Are DeepSeek models free?

Downloadable weights can be obtained without a per-token DeepSeek API fee under their licenses. Running them still costs hardware, power and engineering. The official hosted API is billed by tokens, while the consumer chat service may be used without per-token billing subject to limits.

Are the full V4 models under the MIT license?

The official V4 Flash 0731 model page displays an MIT license. Always inspect the exact model card and dependencies before deployment; different repositories and distilled base models can carry additional terms.

The bottom line

DeepSeek’s model story is no longer “V3 versus R1.” The current hosted platform is V4, with V4 Flash as the economical, agent-focused default and V4 Pro as the higher-capability escalation path. R1 and V3 remain strategically important as downloadable, research and legacy checkpoints, while R1 distills provide the most practical route to local use.

Keep the layers separate: API aliases can change, app/web models may update on a different schedule, and a downloadable checkpoint is fixed. Choose a model through representative evaluation, record the exact version and route harder tasks only when the additional capability justifies its cost.

byDavid Porter

China AI tools DeepSeek Artificial Intelligence

DeepSeek AI: Complete Guide to Models, API and Local Use

How Nvidia Makes Money From AI: Revenue and Business Model

Write a comment