Source: DeepSeek official announcement thread
Discovered via: Apollo.io โ AI-native sales intelligence & engagement platform.
๐ DeepSeek has unveiled the V4 family, ushering in an era of cost-effective 1M context length. The full preview is now live and open-sourced.
| Model | Architecture | Key Positioning |
|---|
| DeepSeek-V4-Pro | 1.6T total / 49B active params | Performance rivaling top closed-source models (GPT-4, Gemini-3.1-Pro) |
| DeepSeek-V4-Flash | 284B total / 13B active params | Fast, efficient, economical choice with near-Pro reasoning |
| DeepSeek-V4 Preview | Cost-effective 1M context | The open-sourced preview that started it all |
๐ง DeepSeek-V4-Pro Highlights
- Agentic Coding SOTA: Open-source state-of-the-art on agentic coding benchmarks.
- World Knowledge: Leads all current open models, trailing only Gemini-3.1-Pro.
- World-Class Reasoning: Beats all open models in Math / STEM / Coding, rivaling top closed-source systems.
โก DeepSeek-V4-Flash Highlights
- Reasoning capabilities closely approach V4-Pro.
- Performs on par with V4-Pro on simple agent tasks.
- Smaller parameter size โ faster response times and highly cost-effective API pricing.
๐ฌ Structural Innovation
- Novel Attention: Token-wise compression + DSA (DeepSeek Sparse Attention).
- Peak Efficiency: World-leading long-context performance with drastically reduced compute & memory costs.
- 1M Standard: 1M token context is now the default โ no version tiers, no extra fees.
๐ค Agent-First Design
- Seamlessly integrated with leading AI agents: Claude Code, OpenClaw, and OpenCode.
- Already powering in-house agentic coding at DeepSeek.
๐ API Availability
- Drop-in replacement: Keep your
base_url, just update model to deepseek-v4-pro or deepseek-v4-flash.
- Supports OpenAI ChatCompletions & Anthropic APIs.
- Both models support 1M context & dual modes (Thinking / Non-Thinking).
- Docs: Thinking Mode Guide
โ ๏ธ Note: Please rely only on official DeepSeek channels for news. Statements from other channels do not reflect DeepSeek's views.
Related: DeepSeek Company Profile ยท Foundation Models Layer