deepseek/deepseek-v4-pro
deepseek/deepseek-v4-pro
1,048,576 context · $1.84/M input tokens · $3.66/M output tokens
DeepSeek V4 Pro is DeepSeek's flagship open-source model released in April 2026, featuring a 1.6T-parameter Mixture-of-Experts architecture with 49B parameters active per token. It supports a 1M-token context window through a novel hybrid attention mechanism combining Compressed Sparse Attention and DeepSeek Sparse Attention, reducing inference FLOPs to 27% and KV cache to 10% compared to V3.2 at million-token scale. Pre-trained on 33T tokens with post-training via GRPO reinforcement learning and on-policy distillation, V4 Pro delivers frontier-level performance in coding (LiveCodeBench 93.5, Codeforces 3206), math (IMOAnswerBench 89.8), and agentic tasks (SWE-bench Verified 80.6) — competitive with GPT-5.4 and Claude Opus 4.6 at a fraction of the cost. It natively supports thinking and non-thinking modes with configurable reasoning effort, function calling, JSON output, and has been specifically optimized for mainstream agent frameworks including Claude Code, OpenClaw, and OpenCode.
Tarification
Paiement à l'usage
Aucun coût initial, payez uniquement ce que vous utilisez
Utilisation de l'API
Utilisez les exemples de code suivants pour intégrer notre API :
from openai import OpenAI
client = OpenAI(
api_key="YOUR_API_KEY",
base_url="https://llm.wavespeed.ai/v1"
)
response = client.chat.completions.create(
model="deepseek/deepseek-v4-pro",
messages=[
{"role": "user", "content": "Hello!"}
]
)
print(response.choices[0].message.content)Introduction au modèle
Deepseek deepseek-v4-pro
DeepSeek-V4-Pro is DeepSeek's most powerful open-source model, released on April 24, 2026. It is a 1.6 trillion parameter Mixture-of-Experts (MoE) language model with 49B active parameters, pre-trained on 33T tokens, supporting a context length of one million tokens. V4-Pro achieves performance on par with top closed-source models like GPT-5.4 and Claude Opus 4.6 across coding, reasoning, and agentic benchmarks — at a fraction of the cost.
Why It Looks Great
- Mixture-of-Experts architecture with 1.6T total parameters and only 49B active for efficient inference
- 1000000 context window powered by Compressed Sparse Attention (CSA) and DeepSeek Sparse Attention (DSA)
- World-class agentic capabilities — optimized for Claude Code, OpenClaw, OpenCode, and CodeBuddy
Key Features
- Context Window: 1000000 tokens
- Max Output: 384000 tokens
- Vision: Not Supported
- Function Calling: Supported
- Thinking Mode: Supported (non-thinking / high / max)
- JSON Output: Supported
- FIM Completion: Supported (non-thinking mode only)
Benchmarks
| Benchmark | V4-Pro | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Verified | 80.6 | 80.8 | — | 80.6 |
| LiveCodeBench | 93.5 | 88.8 | 91.7 | 91.7 |
| Codeforces Rating | 3206 | — | 3168 | 3052 |
| MMLU-Pro | 87.5 | 89.1 | 87.5 | 91.0 |
| IMOAnswerBench | 89.8 | 75.3 | 91.4 | 81.0 |
| Terminal Bench 2.0 | 67.9 | 65.4 | 75.1 | 68.5 |
| Toolathlon | 51.8 | 47.2 | 54.6 | 48.8 |
| BrowseComp | 83.4 | 83.7 | — | 85.9 |
Specifications
| Specification | Value |
|---|---|
| Provider | Deepseek |
| Model Type | Large Language Model (LLM) |
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 1.6T (49B active) |
| Context Window | 1000000 tokens |
| Max Output | 384000 tokens |
| Input | Text |
| Output | Text |
| Vision | Not Supported |
| Function Calling | Supported |
| Thinking Mode | Supported (high / max) |
| Release Date | April 24, 2026 |
How to Use
- Write your prompt — describe the task, provide context, and specify desired output format.
- Submit — the model processes your request and returns the response.
API Integration
Base URL: https://llm.wavespeed.ai/v1 API Endpoint: chat/completions Model ID: deepseek/deepseek-v4-pro
API Usage
Python SDK
from openai import OpenAI
client = OpenAI(
api_key="YOUR_API_KEY",
base_url="https://llm.wavespeed.ai/v1"
)
response = client.chat.completions.create(
model="deepseek/deepseek-v4-pro",
messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)cURL
curl https://llm.wavespeed.ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "deepseek/deepseek-v4-pro",
"messages": [{"role": "user", "content": "Hello!"}]
}'Notes
- Model: deepseek/deepseek-v4-pro
- Provider: Deepseek
- Open-source weights available on HuggingFace and ModelScope
- Supports both OpenAI and Anthropic API formats
- For complex Agent scenarios, use thinking mode with reasoning_effort set to max
Info
Fonctionnalités prises en charge
Guide d'accès API
deepseek/deepseek-v4-pro
DeepSeek V4 Pro is DeepSeek's flagship open-source model released in April 2026, featuring a 1.6T-parameter Mixture-of-Experts architecture with 49B parameters active per token. It supports a 1M-token context window through a novel hybrid attention mechanism combining Compressed Sparse Attention and DeepSeek Sparse Attention, reducing inference FLOPs to 27% and KV cache to 10% compared to V3.2 at million-token scale. Pre-trained on 33T tokens with post-training via GRPO reinforcement learning and on-policy distillation, V4 Pro delivers frontier-level performance in coding (LiveCodeBench 93.5, Codeforces 3206), math (IMOAnswerBench 89.8), and agentic tasks (SWE-bench Verified 80.6) — competitive with GPT-5.4 and Claude Opus 4.6 at a fraction of the cost. It natively supports thinking and non-thinking modes with configurable reasoning effort, function calling, JSON output, and has been specifically optimized for mainstream agent frameworks including Claude Code, OpenClaw, and OpenCode.
Input
$1.84 /M
Output
$3.66 /M
Context
1049K
Max Output
384K
Tool Use
Supported
Try deepseek/deepseek-v4-pro on WaveSpeedAI
Access deepseek/deepseek-v4-pro through our unified API — OpenAI-compatible, no cold starts, transparent pricing.
Open Playground