DeepSeek V4 is officially released and open source: 1.6T parameters, 1M context, FP4 accuracy, fully benchmarked against top closed-source models
DeepSeek officially released the V4 preview version on April 24, 2026, simultaneously opening up the 1.6T parameter V4-Pro (49B activation) and the 284B parameter V4-Flash (13B activation). The new architecture introduces hybrid sparse attention (Hybrid Sparse Attention), manifold constrained hyperconnection (mHC) and FP4 expert weights, supports 1M context length by default, reaches open source SOTA in multiple benchmarks such as Agentic Coding, Math/STEM, etc., and the API price is as low as 75% discount for a limited time.
Core conclusion
The release of DeepSeek V4 marks the first time that an open source large model has truly matched the top closed source models in terms of parameter scale, architectural innovation, and practicality. For AI money-making and automation practitioners, this means: API call costs have dropped significantly (V4-Pro limited time $0.87/M output token), 1M context default support has completely changed the workflow of long document processing, and open source weights can be deployed privately to further lower the compliance threshold. SGLang and Miles have provided inference and RL training support on the first day of release, and their ecological readiness far exceeds any previous generation of DeepSeek models.
Key Points
- Time of incident: 2026-04-24 -Affected objects: AI developers, automated pipeline builders, content production teams
- Core changes: For the first time, the open source model is close to the GPT-5/Gemini-3 level in agentic coding and reasoning capabilities, and the cost is only 1/5 to 1/10 of the closed source model.
Background and release content
On April 24, 2026, DeepSeek officially released the V4 preview version through its API documentation station, and simultaneously open sourced all weights on Hugging Face. This is DeepSeek’s largest cross-generation upgrade after V3, V3.1, R1, and V3.2. It received more than 2,000 points of popularity on Hacker News and became the most watched AI news of the week.
This release contains two model versions:
- DeepSeek-V4-Pro: 1.6T total parameters / 49B activated parameters. Target benchmarking GPT-5, Gemini-3.1-Pro and other top models.
- DeepSeek-V4-Flash: 284B total parameters / 13B activated parameters. It focuses on low-latency and low-cost scenarios and is suitable for production deployment.
Both models support 1M context length and provide both Thinking (reasoning) and Non-Thinking (fast) modes.
Architecture innovation: three core technologies
Compared with the previous generation V3.2, the breakthroughs of DeepSeek V4 are mainly reflected in three key innovations:
| Dimensions | Technology | What it means to us |
|---|---|---|
| Attention mechanism | Hybrid Sparse Attention: Each layer mixes sliding window attention and two compression mechanisms (4:1 top-k or 128:1 dense) | 1M context no longer requires huge computing resources, and the availability of long document processing is greatly improved |
| Residual connections | mHC (Manifold-Constrained Hyper-Connections): manifold-constrained hyper-connections, replacing traditional residual connections | Deep network gradient flow is better, training stability and inference quality are simultaneously improved |
| Quantization accuracy | FP4 expert weight: native FP4 quantization MoE expert layer, adapted to the Blackwell architecture | Maintaining high accuracy while increasing inference speed, more suitable for the latest hardware |
Benchmark performance
DeepSeek V4-Pro’s performance in multiple authoritative reviews:
| Benchmark | Score | Ranking |
|---|---|---|
| GPQA Diamond | 90.1% | Open Source SOTA |
| GSM8K | 92.6% | Open Source SOTA, #3 Global |
| MMLU-Pro | 87.5% | Open Source SOTA |
| HLE | 37.7% | Open source leads |
Former OpenAI researcher Sebastien Bubeck has also publicly commented on DeepSeek V4's capabilities (see related discussion).
API Pricing: Extremely Competitive Pricing Strategy
The pricing plan of DeepSeek V4 is one of the most cost-effective options in the current market:
| model | input (cache hit) | input (cache miss) | output |
|---|---|---|---|
| V4-Flash | $0.028/M | $0.14/M | $0.28/M |
| V4-Pro (limited time 75% off) | $0.03625/M | $0.435/M | $0.87/M |
| V4-Pro (standard price) | $0.145/M | $1.74/M | $3.48/M |
When does the V4-Pro limited time discount last? The official deadline has not been announced, but it is recommended to take advantage as early as possible.
Both models support 384K maximum output tokens, JSON Output, Tool Calls, Chat Prefix Completion and FIM Completion. The API is compatible with OpenAI format (https://api.deepseek.com) and Anthropic format (https://api.deepseek.com/anthropic).
Note: The legacy deepseek-chat and deepseek-reasoner endpoints will be completely retired after July 24, 2026 and are currently automatically routed to deepseek-v4-flash.
Ecological support: SGLang + Miles launched on the first day
On the day of release, the SGLang team and the Miles team simultaneously launched a complete Day-0 support solution:
- Inference optimization: ShadowRadix prefix cache, HiSparse CPU extension KV, MTP speculative decoding, Flash Compressor, Lightning TopK
- Kernel Integration: FlashMLA, FlashInfer, DeepGEMM Mega MoE, TileLang mHC
- RL training: complete parallel solution (DP/TP/SP/EP/PP/CP), supports FP8 training
- Hardware Support: Hopper, Blackwell, Grace Blackwell, AMD, NPU
# SGLang DeepSeek V4
pip install sglang[all]
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V4-Pro --tp 8AI
- ****V4-Flash $0.28/M AI Agent ， GPT-4o/Claude Sonnet 5-10
- ****1M 、、 AI ， RAG
- ****V4-Flash（284B/13B ），
- **Agent ** Agentic Coding SOTA， AI Agent （ n8n、OpenClaw、Claude Code） DeepSeek V4
DeepSeek、OpenAI、ChatGPT、Claude Code、n8n、Gemini、Hermes Agent、OpenClaw、SGLang
Internal link guidance
- Want to learn how? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
- Real case: OpenClaw + Claude Automated Publishing: $1,500–$2,500/mo Case Study
- Realization reference: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo
Topic hub
AI Coding Tools Hub (2026)
From Copilot pricing changes to Claude Code + DeepSeek cost-saving setups—one place to compare tools, read explainers, and follow tutorials.
Explore AI Coding Tools Hub (2026) →Monetization angle
How can you make money from this trend?
WayToClawEarn focuses on verified earn playbooks—not just news. Start from these cases.
DeepSeek + Claude Code Micro SaaS
Run multiple small products on cheap inference
Claude Code bug bounty
Productize agent skills into security services