DeepSeek V4 Pro Lands with 1.6T Parameters, 1M Context, and a Coding Edge
DeepSeek V4 Pro has finally arrived. This is the higher-spec model in the V4 series, using a Mixture-of-Experts (MoE) architecture with a total of 1.6T parameters, 49B activated parameters, support for a 1M (million Token) context length, and a maximum output length of 384K Tokens.
Compared to DeepSeek V4 Flash (284B total parameters / 13B activated parameters), V4 Pro has significantly improved parameter scale, knowledge density, and complex task processing capabilities. With this, the V4 series product structure is now clear: Flash focuses on lower cost and higher concurrency, while Pro provides higher-spec model services for scenarios demanding greater model capability, such as Agents and Coding.
Three Key Architectural Upgrades
Hybrid Attention Mechanism (CSA + HCA)
The V4 series designs a hybrid attention architecture, combining Compressed Sparse Attention and Heavily Compressed Attention. In a 1M Token context scenario, V4 Pro's single-token inference FLOPs are only 27% of V3.2's, and KV Cache usage is only 10%.
This data indicates that when processing ultra-long contexts, the computation required per generated token drops significantly, while also reducing memory usage and deployment costs. For Agent tasks and long-document processing scenarios that need to work within a million-token context, this efficiency improvement can turn million-token context from a theoretical metric into practical productivity.
Manifold-Constrained Hyper-Connections (mHC)
mHC is used to replace traditional residual connections, enhancing the stability of signal propagation between layers while maintaining the model's expressive power. A model with 1.6T parameters faces more severe gradient propagation and stability challenges during training, and mHC provides a structural guarantee for the training stability of large-scale MoE models.
Muon Optimizer
The V4 series adopts the Muon optimizer instead of traditional optimizers, which improves both convergence speed and training stability. With over 32T+ Tokens of pre-training data, improvements at the optimizer level have a significant positive impact on training efficiency.
Training and Post-Training Process
The pre-training phase used over 32T diverse, high-quality Tokens. Post-training adopts a two-stage paradigm:
Independent Domain Expert Cultivation: Expert capabilities are trained separately in different domains through SFT and RL (GRPO).
Unified Model Integration: Expert capabilities from various domains are integrated into a single unified model through On-Policy distillation.
This post-training strategy allows the model to receive targeted optimization in each specialized domain while avoiding the capability conflicts common in multi-task joint training.
Three Inference Modes
Both DeepSeek V4 Pro and V4 Flash support three inference modes, allowing developers to choose based on task complexity.
| Mode | Characteristics | Applicable Scenarios | Response Format |
|---|---|---|---|
| Non-Think | Fast, intuitive responses, no explicit reasoning | Daily conversations, low-risk decisions, routine tasks | response summary |
| Think High | Logical analytical reasoning, slower but more accurate | Complex problem-solving, planning tasks | thinking thinking response summary |
| Think Max | Pushes reasoning to the limit | Competition-level math, extreme programming, exploring model capability boundaries | Special system prompt + thinking thinking response summary |
In Think Max mode, it is recommended to set the context window to at least 384K Tokens to ensure the model has sufficient space for deep reasoning.
Benchmark Evaluation Data: V4-Pro Max vs. Top-Tier Models
Knowledge and Reasoning
| Benchmark | Opus-4.6 Max | GPT-5.4 xHigh | Gemini-3.1-Pro High | DS-V4-Pro Max |
|---|---|---|---|---|
| MMLU-Pro | 89.1 | 87.5 | 91.0 | 87.5 |
| SimpleQA-Verified | 46.2 | 45.3 | 75.6 | 57.9 |
| Chinese-SimpleQA | 76.4 | 76.8 | 85.9 | 84.4 |
| GPQA Diamond | 91.3 | 93.0 | 94.3 | 90.1 |
| HLE | 40.0 | 39.8 | 44.4 | 37.7 |
On knowledge tasks, V4-Pro Max performs well on SimpleQA-Verified (57.9) and Chinese-SimpleQA (84.4). The former surpasses Opus-4.6 Max (46.2) and GPT-5.4 xHigh (45.3), and the Chinese knowledge Q&A score also indicates that V4 Pro's training quality on Chinese corpora is solid.
Coding and Math
| Benchmark | Opus-4.6 Max | GPT-5.4 xHigh | Gemini-3.1-Pro High | DS-V4-Pro Max |
|---|---|---|---|---|
| LiveCodeBench | 88.8 | - | 91.7 | 93.5 |
| Codeforces Rating | - | 3168 | 3052 | 3206 |
| HMMT 2026 Feb | 96.2 | 97.7 | 94.7 | 95.2 |
| IMOAnswerBench | 75.3 | 91.4 | 81.0 | 89.8 |
| Apex Shortlist | 85.9 | 78.1 | 89.1 | 90.2 |
Coding is an area where V4-Pro Max performs strongly, with good results on LiveCodeBench 93.5 and Codeforces Rating 3206. In math, HMMT and IMO scores are competitive with GPT-5.4 xHigh, also at a high level.
Agent Tasks
| Benchmark | Opus-4.6 Max | GPT-5.4 xHigh | Gemini-3.1-Pro High | DS-V4-Pro Max |
|---|---|---|---|---|
| SWE Verified | 80.8 | - | 80.6 | 80.6 |
| SWE Pro | 57.3 | 57.7 | 54.2 | 55.4 |
| BrowseComp | 83.7 | 82.7 | 85.9 | 83.4 |
| MCPAtlas Public | 73.8 | 67.2 | 69.2 | 73.6 |
| Toolathlon | 47.2 | 54.6 | 48.8 | 51.8 |
On Agent tasks, V4-Pro Max is in the mainstream tier. The SWE Verified 80.6% solve rate and MCPAtlas 73.6 score indicate that V4 Pro has practical utility in real software engineering and tool-calling scenarios.
API Feature Coverage
The API features provided by DeepSeek V4 Pro support mainstream Agent development needs:
JSON Output: Structured output for easy integration with downstream systems.
Tool Calls / Function Calling: Tool calling capability.
Responses API: Standard response interface.
Anthropic API Compatibility: Can be called directly using the Anthropic format.
Conversation Prefix Continuation (Beta): Supports continuing generation from a specified prefix.
FIM Completion (Beta, Non-Think mode only): Fill-in-the-Middle code completion.
Access methods are compatible with both OpenAI and Anthropic protocol formats. For projects with existing integration interfaces, migration can be done by directly replacing the Base URL and API Key.
Pricing and Cost Analysis
| Billing Item | DeepSeek V4 Flash | DeepSeek V4 Pro | Pro / Flash Multiple |
|---|---|---|---|
| Cache Hit Input | 0.02 CNY / M Tokens | 0.025 CNY / M Tokens | 1.25x |
| Cache Miss Input | 1 CNY / M Tokens | 3 CNY / M Tokens | 3x |
| Output | 2 CNY / M Tokens | 6 CNY / M Tokens | 3x |
| Concurrency Limit | 2500 | 500 | - |
Here are the details to note:
Low Cache Hit Price: V4 Pro cache hit input is only 0.025 CNY / M Tokens. For Agent tasks with a large amount of repetitive prefixes (like fixed system prompts + tool descriptions), reasonable use of caching can control input costs.
Output Cost is the Main Expense: V4 Pro output is 6 CNY / M Tokens. In Think Max mode, the reasoning process itself consumes more output Tokens.
Significant Difference in Concurrency Limits: V4 Flash's concurrency limit is 5 times that of V4 Pro, making it suitable for high-frequency calling scenarios.
Flash vs. Pro Selection Guide
Daily Conversations, Lightweight Tasks: Recommend V4 Flash (Non-Think / Think High), fast response and low cost.
Complex Code Generation, Code Review: Recommend V4 Pro (Think High), suitable for scenarios requiring high code comprehension.
Multi-step Agent Tasks, Software Engineering: Recommend V4 Pro (Think High / Think Max), supports long-process tasks.
Batch Data Processing and Extraction: Recommend V4 Flash (Non-Think), leveraging its high concurrency and low cost advantages.
API Engineering Management in a Multi-Model Environment
With the addition of DeepSeek V4 Pro, the number of high-quality AI models development teams can access has increased. When using multiple APIs like OpenAI, Anthropic, Gemini, and self-built relays simultaneously, adapting to different protocol formats, channel switching, traffic distribution, and multi-project virtual Key management brings additional engineering overhead.
ServBay's AI Gateway (AI Gateway) provides a localized unified entry management solution:
- Multi-Channel Access and Distribution: Supports adding various official APIs, subscription accounts, and relays, enabling automatic traffic distribution and channel priority configuration.
- Hot Swap and Automatic Switching: Supports manual switching without service interruption or automatic switching upon channel failure.
- Model Mapping: Allows mapping model names in code (e.g.,
claude-opus-5) to the actual called model (e.g.,glm-5.2), without modifying downstream application code. - Protocol Conversion: Shields protocol differences between OpenAI, Anthropic, Gemini, etc. Applications only need to interface with the gateway's unified interface, and the gateway is responsible for automatically converting requests to the target service provider's format.
- Virtual Keys: Create independent local virtual Keys for different development projects, tracking their respective Token consumption and call volume for easier internal cost accounting.
Through such local management tools, developers can more conveniently route between DeepSeek V4 Pro and other models, reducing the repetitive adaptation work caused by protocol and channel changes.