Overview
DeepSeek V4 is the new generation flagship MoE large model officially released by DeepSeek on 2026/4/24, open-sourced simultaneously on Hugging Face and ModelScope, accompanied by a 58-page technical report. Two tiers are released at once: **V4-Pro** (total parameters 1.6T / activated 49B / 384 routing experts + 1 shared) and **V4-Flash** (total parameters 284B / activated 13B / 256 routing experts).
**Native 1M token context for all tiers** (not exclusive to paid tiers), significantly compressing long-context computation and KV cache through **DSA (DeepSeek Sparse Attention)** sparse attention. The API is compatible with both OpenAI ChatCompletions and Anthropic specifications, natively adapting to mainstream Agent toolchains such as Claude Code, OpenClaw, OpenCode, and CodeBuddy.
**Key Milestone Timeline**:
- **2026/4/24**: Open-source release + Ascend zero-day adaptation
- **2026/5/22**: V4-Pro permanent price reduction of 75% (input ¥3.1/M, output ¥6.2/M), approximately 98% cheaper than GPT-5.5; concurrently established the **Harness team** (led personally by Liang Wenfeng) to tackle Code agents competing with Claude Code
- **2026/6/16**: Completed the largest single-round financing in Chinese AI history of **~$7 billion (approximately ¥50 billion)**, post-investment valuation of $50 billion (Tencent $10 billion + CATL $5 billion + Liang Wenfeng personal $20 billion)
- **2026/6/20**: World's first to run on Huawei Ascend AI computing chips—deploying cutting-edge models on purely domestic computing power is no longer a PPT promise
- **2026/6/4**: A landmark event for DeepSeek's overseas expansion: US top AI Agent platform **[Lindy](/tools/lindy.html)** (backed by Sequoia, 500,000 active users) CEO Flo Crivello officially announced on X that **100% of production traffic was migrated from Anthropic Claude to DeepSeek V4**, taking 9 months to complete the technical switch, saving the company millions of dollars annually. This is DeepSeek's first large-scale case of a 'top Western Agent platform completely abandoning Claude for DeepSeek', and a strong endorsement of DeepSeek V4's production-grade capabilities in the international market—Crivello stated 'using Claude is more expensive than paying employee salaries', and after the switch, Agent task performance was 'on par with or even better than Claude'.
- **2026/7/15**: V4 official version fully launched: Pro / Flash dual architecture, **pioneering peak-valley time-of-use pricing + cache hit discounts**—API call prices double during peak computing hours on workdays (9:00-12:00, 14:00-18:00), while maintaining ultra-low inclusive pricing during off-peak hours and all weekends; cache hits further reduce costs by 90%, and off-peak scheduling can halve call costs, breaking the domestic large model 'single low-price involution' pricing paradigm
- **2026/7/15**: Annualized revenue revealed: According to The Information, DeepSeek's recent annualized recurring revenue (ARR) has reached **$400-500 million**, almost entirely from API calls, with gross margins exceeding 70%; concurrently launched a second round of financing, targeting **¥50 billion**, with post-investment valuation rising to **¥500 billion (approximately $70 billion)**, a significant jump from the first round valuation of ¥50 billion a month earlier; **has initiated IPO preparation on the STAR Market, aiming to submit an application within the year and list in 2027**
LiveCodeBench 93.5, Codeforces 3206 (human rank 23rd), Putnam-2025 perfect score 120/120—it is the current ceiling for open-source coding and mathematics.
Key Features
- 1.6T MoE + Native 1M Context: Pro total params 1.6T / activated 49B, Flash 284B / activated 13B, all tiers natively support 1M context at no extra cost
- DSA Sparse Attention: For 1M context scenarios, Pro computation reduced to 27%, KV cache compressed 9.5×; Flash compressed 13.7×
- Thinking / Non-Thinking Dual Mode: Supports reasoning_effort=high/max switching, HLE improves from 7.7 (non-thinking) to 37.7 (max)
- Native Agent Framework Adaptation: Out-of-the-box integration with Claude Code, OpenClaw, OpenCode, CodeBuddy; API compatible with both OpenAI and Anthropic interfaces
- Open-Source Coding Ceiling: LiveCodeBench 93.5, Codeforces 3206 (human rank 23rd), Putnam-2025 perfect score 120/120
- World's First to Run on Ascend: On 2026/6/20, world's first to run on Huawei Ascend AI computing chips, making deployment of cutting-edge open-source models on purely domestic computing power a reality
- Harness Team Focused on Code Agent: Since 5/22, Liang Wenfeng personally leads the Harness team to tackle Code agents competing with Claude Code
- Permanent 75% Price Reduction + Peak-Valley Time-of-Use Pricing (Launched): Since 5/22, V4-Pro permanent price reduction of 75% (98% cheaper than GPT-5.5); on 7/15, official version fully launched peak-valley time-of-use pricing—prices double during peak workday hours (9-12, 14-18), while maintaining inclusive pricing during off-peak hours and weekends; off-peak scheduling can save more than half the cost
Use Cases
- Enterprises requiring private deployment of cutting-edge models (including purely domestic computing power stacks)
- API call scenarios with tight budgets and pursuit of extreme cost-effectiveness
- High-difficulty tasks such as code generation, math competitions, and long codebase understanding
- Long document / full codebase analysis (native 1M context)
- Replacing Claude Code / CodeBuddy backends to reduce costs
- Model selection for state-owned enterprises / Xinchuang scenarios (full Ascend stack adaptation)
Pros
- Pro output price ¥6.2/M token (after permanent reduction), approximately 1/50 of GPT-5.5
- Open-source ceiling in coding and mathematics, partially surpassing closed-source flagships
- Native 1M context standard across all tiers, not exclusive to paid tiers
- API compatible with both OpenAI and Anthropic interfaces, extremely low migration cost
- World's first to run on Huawei Ascend, fully domestic computing power stack enabled
- Weights fully open-source, commercially usable and privately deployable
- Cache hit input price as low as ¥0.2~¥1 / M token
Pricing
**V4-Pro** (after permanent price reduction on 5/22): Input ¥3.1/M (cache hit ¥0.31), output ¥6.2/M. **V4-Flash**: Input ¥1/M (cache hit ¥0.2), output ¥2/M. **On 2026/7/15, official version fully launched peak-valley time-of-use pricing**: API call prices double during peak computing hours on workdays (9:00-12:00, 14:00-18:00), while maintaining ultra-low inclusive pricing during the remaining 17 off-peak hours + all weekends; combined with cache hit discounts, off-peak scheduling can halve call costs. chat.deepseek.com and the official app remain free to use. **Ascend private deployment**: Since 6/20, official support for batch deployment on Ascend 950 clusters, priced per node.
Summary
DeepSeek V4 completed a six-step journey in the first half of 2026: 4/24 open-source release + Ascend zero-day adaptation → 5/22 permanent 75% price reduction + establishment of Harness team for Code Agent → 6/16 completed ~$7 billion financing with valuation exceeding $50 billion → 6/20 world's first to run on Huawei Ascend → 7/15 official version fully launched with peak-valley time-of-use pricing → **7/15 annualized revenue revealed at $400-500 million, IPO preparation on STAR Market initiated, valuation target soaring to ¥500 billion**. This combination integrates 'cutting-edge capabilities + extreme cost-effectiveness + purely domestic computing power + flexible pricing + commercial validation', making it the most representative product of China's full-stack domestic AI in 2026. If you need a purely domestic stack for enterprise AI, or simply want the most cost-effective cutting-edge model via API, DeepSeek V4 has almost no comparable alternative; teams scheduling tasks off-peak (avoiding workday 9-12 and 14-18) using peak-valley pricing can save more than half the cost. For long Agent tasks and general knowledge, it can still be complemented with Claude Code / GPT-5.6.
Version History
- Annualized revenue revealed + STAR Market IPO preparation (2026-07-15): The Information reports ARR of $400-500 million, gross margin over 70%; second financing round launched targeting ¥50 billion, valuation rising to ¥500 billion; IPO preparation on STAR Market initiated, aiming to submit application within the year
- V4 official version fully launched · Peak-valley time-of-use pricing (2026-07-15): Pro / Flash dual architecture official version fully launched; pioneered peak-valley time-of-use pricing (workday 9-12/14-18 peak double, rest off-peak + weekend inclusive) + cache hit discounts, breaking domestic large model single low-price involution paradigm
- Lindy switches to DeepSeek · DeepSeek overseas landmark event (2026-06-04): US top AI Agent platform Lindy announces 100% production traffic migration from Anthropic Claude to DeepSeek V4, completed over 9 months, saving millions annually; first case of a top Western Agent platform completely abandoning Claude for DeepSeek
- V4 world's first to run on Ascend (2026-06-20): World's first to run on Huawei Ascend AI computing chips, deploying cutting-edge models on purely domestic computing power, marking a milestone in China's full-stack AI domestic production
- $50 billion valuation financing (2026-06-16): Completed ~$7 billion first-round financing, post-investment valuation $50 billion; Tencent $10 billion + CATL $5 billion + Liang Wenfeng personal $20 billion
- V4-Pro permanent 75% price reduction + Harness team (2026-05-22): Pro permanent 75% price reduction (98% cheaper than GPT-5.5); Liang Wenfeng leads establishment of Harness team competing with Claude Code
- V4 open-source release (2026-04-24): 1.6T MoE + 1M context + Ascend zero-day adaptation, coding and math open-source ceiling, 58-page technical report