DeepSeek V4 Pro

What is DeepSeek V4 Pro?#

DeepSeek V4 Pro is DeepSeek's largest language model, a 1.6 trillion parameter mixture-of-experts with 49B parameters active per token. It first appeared as a preview on April 24, 2026, and the general availability version, DeepSeek-V4-Pro-0813, arrived on August 13, 2026 with much stronger agent results. On Unifically it runs as deepseek/deepseek-v4-pro with 1M tokens of context, up to 384K output tokens, and thinking on by default. It takes text only. Since September 10, 2026 the smaller DeepSeek V4.1 Flash beats it on most benchmarks, so V4 Pro now makes sense mainly for work that depends on broad world knowledge.

Key features of DeepSeek V4 Pro#

Independent score: 36 on the Intelligence Index

Artificial Analysis scores V4 Pro 0813 at 36 on its Intelligence Index at max effort, above V4 Flash 0731 at 34 and below V4.1 Flash at 39, as of September 23, 2026. It measures output at about 66 tokens per second, slower than average for its class, and less verbose than the Flash models: 160M output tokens to run the index against 250M for V4.1 Flash.

Agent gains in the 0813 release

DeepSeek reports 87.9% on Terminal-Bench 2.1, 62.7% on DeepSWE, 74.1% on Toolathlon-Verified, and 83.3% on CyberGym for the GA version. Humanity's Last Exam scores 42.7% without tools and 60.0% with them. These are DeepSeek's own results.

The most world knowledge in the DeepSeek lineup

In DeepSeek's base-model table, V4 Pro scores 55.2% on SimpleQA-Verified against 42.3% for V4.1 Flash, and 51.5% against 45.2% on LongBench-V2. Facts recalled without a search tool and long-document reasoning are where the extra size still shows.

1M context with sparse attention

The V4 family pairs token-wise compression with DeepSeek Sparse Attention, which is what makes a 1M token window affordable to serve. A single response can run to 384K tokens, enough for a full report or a large generated file.

Best for#

Use cases#

Build a research assistant that answers domain questions from what the model knows and cites the documents you pass in. Run a document review pipeline that reads a full contract set or a filing pack in one request and returns a structured summary. Use it as the planner in a two-model agent, with V4 Pro deciding what to do and a faster model carrying out the routine steps. Teams already on an OpenAI-compatible client can call it with a model ID change.

Limitations#

V4.1 Flash is ahead on most of DeepSeek's own benchmarks: 74.2% against 62.7% on DeepSWE v1.1 and 54.8% against 43.2% on AutomationBench. On the harder Terminal-Bench 3.0 and 4.0 suites, V4 Pro scores 11.8% and 12.4%.

Speed is a real cost. At about 66 output tokens per second, a long thinking response takes several times longer than on V4.1 Flash.

Human preference votes are middling. On Arena's Text board, deepseek-v4-pro-high-20260813 sits at #50 with 1463 Elo on 9,008 votes, as of September 23, 2026.

It does not accept images. DeepSeek has also said a V4.1 Pro will follow, and it once announced routing V4 Pro traffic to V4.1 Flash before reversing course, so plan for this model ID to change, with notice. Peak-hour requests bill at 2x the off-peak rate.

DeepSeek V4 Pro vs DeepSeek V4.1 Flash#

V4.1 Flash wins on agent work and speed: DeepSWE v1.1 at 74.2% against 62.7%, Terminal-Bench 2.1 at 90.6% against 87.9%, and more than three times the output speed on Artificial Analysis. V4 Pro wins on knowledge: GPQA Diamond at 92.4% against 90.9%, Humanity's Last Exam at 42.7% against 36.8%, and a clear lead on SimpleQA-Verified in the base-model table. Default to V4.1 Flash and switch to V4 Pro when answers depend on what the model knows rather than what it can do with tools.

When to use DeepSeek V4 Pro#

Use DeepSeek V4 Pro for knowledge-heavy questions, expert reasoning, and long-document analysis where speed is not the constraint. For coding agents and high-volume tool use, DeepSeek V4.1 Flash is faster and scores higher.

API examples#

Call DeepSeek V4 Pro by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Parameter docs live at docs.unifically.com/models/llm/deepseek/deepseek-v4-pro.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "deepseek/deepseek-v4-pro",
    "messages": [
      { "role": "user", "content": "Summarize the key obligations and deadlines in this supply contract." }
    ]
  }'

The response comes back synchronously with the answer in content and the thinking in reasoning_content. Set "stream": true to receive tokens as they generate.