DeepSeek Models Explained: Complete Model Overview

DeepSeek has released general language, reasoning, coding, vision-language, mathematics and research models. Only two model IDs were listed for the first-party API on 2026-07-30: deepseek-v4-flash and deepseek-v4-pro. Older families remain useful for history, research and local deployment but should not be presented as current API products.

Status map

“Historical” does not mean useless. It means a reader should not assume the model is the current hosted default or API alias.

DeepSeek V4 Flash and V4 Pro

DeepSeek announced V4 Preview in April 2026. The official API pricing page lists both models with a one-million-token context window, up to 384,000 output tokens, thinking and non-thinking modes, JSON output, tool calls, prefix completion and non-thinking FIM completion.

DeepSeek positions Flash for speed and economy and Pro for more demanding work. Both are text models in the API. An integration may proxy image descriptions through another model, but that does not make V4 natively multimodal. Verify current prices and model IDs before deployment.

DeepSeek V3 family

DeepSeek V3 is a 671-billion-parameter mixture-of-experts model with 37 billion activated parameters per token according to its official repository. Its model card lists 128K context for the released weights. Later V3.x service updates changed hosted aliases over time.

V3 remains important for understanding DeepSeek’s architecture and the base used for early R1 work. It is not a safe current API name. The legacy deepseek-chat and deepseek-reasoner aliases were scheduled for retirement after 2026-07-24 and should be removed from new examples.

DeepSeek R1

R1 is the first major DeepSeek reasoning family. The official repository lists R1 and R1-Zero at 671B total parameters with 128K context, plus smaller distilled checkpoints based on Qwen and Llama families. R1-0528 later updated the main model and an 8B distilled variant.

Use R1 for research into explicit reasoning workflows or local experiments where an appropriate checkpoint fits. Read licences for the exact distilled base. The R1 guide explains model naming, local access and limitations.

DeepSeek Coder and Coder V2

DeepSeek Coder began as a family from 1B to 33B trained heavily on code. Coder V2 continued from DeepSeek V2 and released 16B and 236B mixture-of-experts variants, with 2.4B and 21B active parameters respectively. Its model card describes a 128K context and support for many programming languages.

Coder V2 is a downloadable specialised family, not a current first-party API ID. Developers deciding between it and current V4 should compare local control, hardware, licence and real repository tests. See the Coder guide.

DeepSeek VL and VL2

DeepSeek VL combined image and text in 1.3B and 7B families. VL2 introduced Tiny, Small and larger mixture-of-experts variants with 1.0B, 2.8B and 4.5B activated parameters. Official materials describe visual question answering, OCR, document, chart and grounding tasks.

These are specialised open model repositories. They should not be confused with the current text-only V4 API or hosted file extraction. The VL guide covers supported workflows and deployment cautions.

Research and domain models

DeepSeek’s official GitHub organisation also contains projects for mathematics, theorem proving, OCR, sparse attention, systems and inference. Some publish weights; others publish code or papers. Their purpose can be narrower than conversational assistance.

Check repository ownership, release date, model card, licence, dependencies and maintenance state. A repository name alone does not establish production support. Research code may require specific CUDA, PyTorch or Linux environments.

Model IDs, versions and checkpoints

An API ID selects a hosted service. A version identifies the underlying release. A checkpoint is a set of downloadable weights. A runtime tag, such as an Ollama model name, is a third party’s packaging. These strings are not interchangeable.

For production, record the API response’s model and system fingerprint where available. For local use, pin the exact checkpoint revision and quantisation. Avoid the bare word “DeepSeek” in evaluation records.

Licences and “open source”

DeepSeek uses different licence arrangements across repositories and derived models. R1’s official repository states MIT terms for its released model weights, while distilled variants depend on Qwen or Llama bases. Older Coder and VL repositories include their own model licences.

Read the files attached to the exact artifact. Do not assume one DeepSeek licence covers every family, derivative or third-party quantisation. Commercial permission does not remove privacy, safety or downstream legal obligations.

How to choose

Use V4 Flash for current low-cost first-party API tasks that pass your evaluation. Use V4 Pro when harder tasks justify its rate. Choose an R1 distilled checkpoint for local reasoning experiments on attainable hardware. Consider Coder V2 for specialised local code research and VL2 for image-text experiments.

The model decision guide translates those options into a task, privacy, budget and hardware checklist.

Conclusion

DeepSeek’s model catalogue is a timeline, not a flat menu. Start with current API status, then distinguish hosted IDs from downloadable checkpoints and research repositories. Version every claim and test the exact model you plan to use.