V4-Pro and V4-Flash, both with 1M context, Codeforces 3206 surpasses Claude Opus 4.6
距 2026-07-24 还有 19 天。deepseek-chat 与 deepseek-reasoner 两个旧模型名停止使用。当前阶段分别指向 V4-Flash 非思考模式与思考模式，迁移只需改 model 参数，base_url 不变。
From V4 release in April 2026 to image recognition and DeepCode, then 51B yuan funding and July 24 API deprecation — DeepSeek's busy half-year.
DeepSeek released V4-Pro (1.6T total / 49B activated) and V4-Flash (284B / 13B), both MoE models with 1M token context standard, MIT licensed. Three architectural breakthroughs: hybrid attention (CSA + HCA), manifold-constrained hyperconnection (mHC), Muon optimizer.
V4-Pro-Max mode (maximum reasoning tier) sets new open-source records: LiveCodeBench 93.5%, Codeforces 3206, IMO-AnswerBench 89.8%, HMMT 2026 95.2%, SWE-bench Verified 80.6%, MRCR 1M long-context retrieval 83.5%. DeepSeek officially stated V4-Pro-Max is "firmly establishing itself as the best open-source model available today".
V4's core architectural breakthrough is hybrid attention: CSA compresses history at 4:1 ratio, HCA compresses ultra-long text at 128:1, SWA tracks the most recent 128 tokens. Combined with mHC and Muon optimizer, V4 uses only 27% of V3.2's FLOPs and 10% of its KV cache at 1M context.
Image recognition went beta a few days after V4 release. Unlike previous OCR-based text extraction, the new V4 natively recognizes image content: scenes, clothing details, team crests on railings, and even detects AI-generated images through lighting, skin texture, and edge smoothness.
DeepSeek launched its official CLI AI coding assistant DeepCode, similar to Claude Code: scans project directory structure, analyzes Controller/Service call relationships, auto-installs dependencies and verifies execution. DeepSeek-model-optimized with lower cost for long conversations.
DeepSeek closed 51B yuan first external funding round (post-money valuation near 400B yuan), with founder Liang Wenfeng personally investing ~20B yuan as the largest single backer. The company doubles headcount across all teams, focusing on AI search, agent infrastructure, and data research.
V4 didn't take the "parameter bloat" route. Architectural innovation made 1M context accessible and Agent capability cost-effective.
V4-Pro-Max achieves multiple SOTA among open-source models across coding, math, and long-context.
| Benchmark | Category | V4-Pro-Max | Notes |
|---|---|---|---|
| LiveCodeBench | Live Coding | 93.5% | New open-source SOTA |
| Codeforces Rating | Competitive Programming | 3206 | Surpasses Gemini-3.1-Pro and Claude Opus-4.6 |
| SWE-bench Verified | Real Software Engineering | 80.6% | Close to Claude Opus 4.6 (80.8%) |
| HumanEval pass@1 | Code Generation | 90.8% | Surpasses Claude 3.5 Sonnet, matches GPT-4o |
| AIME 2026 | Math Competition | 99.4% | Near-perfect |
| IMO-AnswerBench | Math Olympiad | 89.8% | Open-source leading |
| HMMT 2026 | Math Tournament | 95.2% | Open-source leading |
| MMLU-Pro | Multi-subject Reasoning | 87.5% | Open-source leading |
| MRCR 1M | 1M Context Retrieval | 83.5% | Surpasses GPT-5 (69.8%) |
Source: DeepSeek V4 official technical report, April 2026. Benchmark scores may vary with future updates.
编程能力深度评测 | 看长文本应用 → 百万上下文应用 | 跟 GPT-5 / Claude 对比 → V4 vs GPT-5 vs Claude | 7-24 旧 API 迁移 → 迁移指南" data-en="Full coding deep-dive → coding benchmark page | 1M context use cases → long-context page | V4 vs GPT-5 vs Claude → comparison page | July 24 legacy migration → migration guide.">Full coding deep-dive → coding benchmark page | 1M context use cases → long-context page | V4 vs GPT-5 vs Claude → comparison page | July 24 legacy migration → migration guide.
LiveCodeBench 93.5%, Codeforces 3206, HumanEval 90.8%, SWE-bench 80.6%, open-source SOTA. Internal agent coding experience surpasses Claude Sonnet 4.5.
Both V4-Pro and V4-Flash ship with 1M token context standard. Hybrid attention keeps long-text cost manageable (inference FLOPs only 27% of V3.2).
Non-Think / Think High / Think Max three reasoning tiers balance response speed and depth. Use Non-Think for simple Q&A, Think Max for complex agent tasks.
Image recognition went beta after V4 release. Native image recognition (not OCR) identifies scenes, clothing details, and detects AI-generated images.
Pro output ¥24/MTok, Flash output ¥2/MTok, just 1/10 to 1/30 of Claude Opus 4.6 ($75/MTok).
Full series MIT licensed, model weights and technical report both open source. Developers can freely commercialize, modify, deploy. Native support for Ascend, Cambricon domestic chips.
深度对比页。" data-en="Core specification comparison (latest 2026 data). Full comparison + test cases → comparison page.">Core specification comparison (latest 2026 data). Full comparison + test cases → comparison page.
| Dimension | DeepSeek V4-Pro | Claude Opus 4.6 | GPT-5.4 |
|---|---|---|---|
| Context window | 1,000,000 tokens | 200,000 tokens | 400,000 tokens |
| LiveCodeBench | 93.5% | ~88% | ~90% |
| Codeforces Rating | 3206 | ~3000 | 3168 |
| SWE-bench Verified | 80.6% | 80.8% | ~80% |
| MRCR 1M Long-Context | 83.5% | — | 69.8% |
| Output Price/Mtok | ¥24 | $75 | $10-30 |
| License | MIT open source | Closed | Closed |
Source: DeepSeek V4 official report, Anthropic, OpenAI, benchlm.ai, pricepertoken.com — May 2026 data.
July 2025 - January 2026
Liang Wenfeng's paper wins ACL 2025 Best Paper Award; Engram conditional memory module technology revealed early.
February - April 2026
DeepSeek App 1.7.4 expands context to 1M, knowledge base updated to May 2025, gray-scale beta
April 24, 2026
V4-Pro and V4-Flash released together, MIT licensed, 1M context standard
May 2026
Native image recognition went beta, DeepSeek's official CLI coding assistant DeepCode launched.
July 24, 2026 (19 days)
/deepseek-v3-to-v4-migration.html" data-en="deepseek-chat and deepseek-reasoner legacy names deprecated. Only need to change model parameter. Full guide → /deepseek-v3-to-v4-migration.html.">deepseek-chat and deepseek-reasoner legacy names deprecated. Only need to change model parameter. Full guide → /deepseek-v3-to-v4-migration.html.
DeepSeek V4 preview released April 24, 2026. V4-Pro (1.6T total / 49B activated) and V4-Flash (284B / 13B) launched together, MIT licensed, 1M token context standard.
编程能力深度评测页。" data-en="V4-Pro-Max sets new open-source records: LiveCodeBench 93.5%, Codeforces Rating 3206 (surpassing Gemini-3.1-Pro and Claude Opus-4.6), SWE-bench Verified 80.6%, HumanEval pass@1 90.8%, AIME 2026 99.4%. Full coding analysis → coding benchmark page.">V4-Pro-Max sets new open-source records: LiveCodeBench 93.5%, Codeforces Rating 3206 (surpassing Gemini-3.1-Pro and Claude Opus-4.6), SWE-bench Verified 80.6%, HumanEval pass@1 90.8%, AIME 2026 99.4%. Full coding analysis → coding benchmark page.
迁移指南。" data-en="From July 24, 2026, the legacy model names deepseek-chat and deepseek-reasoner are deprecated. They currently map to V4-Flash non-think and think modes. Full migration guide with code examples → migration guide.">From July 24, 2026, the legacy model names deepseek-chat and deepseek-reasoner are deprecated. They currently map to V4-Flash non-think and think modes. Full migration guide with code examples → migration guide.
百万上下文应用页。" data-en="V4 and Gemini 3.1 series are in the same 1M-context tier, but V4 costs only 1/20 of Gemini 3.1; Claude 3.5/4 series has only 200K context. V4 scores 83.5% on MRCR 1M retrieval, surpassing GPT-5's 69.8%. Full long-context test → long-context page.">V4 and Gemini 3.1 series are in the same 1M-context tier, but V4 costs only 1/20 of Gemini 3.1; Claude 3.5/4 series has only 200K context. V4 scores 83.5% on MRCR 1M retrieval, surpassing GPT-5's 69.8%. Full long-context test → long-context page.
V4 series uses MIT license, model weights and technical report both published on Hugging Face. Developers can freely commercialize, modify, deploy. Native support for Ascend, Cambricon domestic chips.
V4 strong points: repo-level code understanding and generation, ultra-long document analysis (financial reports, contracts, novels), agent coding, low-cost batch text processing, domestic compute adaptation. Weak points: high-aesthetic frontend generation, competition-level math/science reasoning, ultra-complex multi-turn conversations (context forgetting after 15 turns), high-precision multimodal creation.