DeepSeek V4 Released: The Most Powerful Open-Source Model! Agent Performance Fully Surpasses Mainstream Closed-Source Models at Just 2% of the Price

After months of rumors, DeepSeek—hailed as the “strongest open-source warrior”—officially announced its latest large language model today (April 24, 2026). DeepSeek V4 Preview officially launched and fully open-sourced. This next-generation model not only breaks through to 1.6 trillion parameters in scale, but also 1M Ultra-Long Context Listed as standard across the entire series. According to DeepSeek’s official disclosure, the V4 series Agent Execution Capabilitiesachieved a qualitative leap, with its actual code delivery quality surpassing the well-known Claude 3.5 Sonnet in internal company testing. And what Jensen Huang fears most has occurred: the training and inference of DeepSeek V4 was entirely completed using Huawei’s domestically produced GPUs.

Dual Version Launch: Pro and Flash Positioning Strategy

This release includes two core models, both supporting native multimodal input (text, images, audio and video):

1. DeepSeek-V4-Pro (Flagship)
   • Total parameters1.6 trillion (1.6T)
   • activation parametersApproximately 49 billion (49B) activations per token
   • PositioningSpecifically designed for complex agent tasks, challenging programming development, and deep reasoning. Its performance has reached the highest level among current open-source models, rivaling world-class top-tier closed-source models.
2. DeepSeek-V4-Flash (High-Speed Version)
   • Total parameters:284 billion (284B)
   • Activation parametersApproximately 13 billion (13B) activations per token
   • PositioningEmphasizing ultra-fast response and cost-effectiveness. On par with the Pro version in simple Agent tasks, making it the top choice for applications requiring high throughput and low cost.

Agent Capability Enhancement: Internal Testing Outperforms Sonnet

DeepSeek V4 this time Agentic Coding They have put in significant effort in this area, adapting and optimizing for mainstream agent products such as Claude Code, OpenClaw, OpenCode, and CodeBuddy, resulting in improved performance across code tasks and documentation generation. According to the company, V4-Pro has become the primary coding model used daily by DeepSeek employees, and it outperforms several well-known mainstream closed-source models across multiple capability metrics.

• Delivery Quality BenchmarkingInternal review feedback shows that V4-Pro’s user experience is better than Claude 3.5 Sonnetand the delivery quality is already very close Claude Opus 4.6 Non-Thinking ModeAlthough it still falls slightly short of Opus 4.6’s “thinking mode” when dealing with extremely complex logic, it is already unmatched in the open-source world.
• Mainstream Framework CompatibilityNew model targeting Claude Code、OpenCode、CodeBuddy、OpenClaw …and other mainstream Agent products have been deeply optimized, significantly improving cross-file code refactoring and coherence in document generation.

Performance Benchmark: World Knowledge and Reasoning Capabilities

• World KnowledgeIn comprehensive world knowledge benchmarks, V4-Pro significantly outperforms other open-source models, ranking just below top closed-source models. Gemini-Pro-3.1。
• STEM and ReasoningIn math, science, and competitive programming benchmarks, V4-Pro achieved performance on par with the world’s top closed-source models, demonstrating strong logical reasoning capabilities.1M Context UniversalDeepSeek announces “1M context will be the standard for all future official services.” Through its innovative Token dimension compression technologyand DSA Sparse Attention MechanismV4 significantly reduces computational and VRAM requirements under a 1 million token load.

Architecture Innovation: Engram and mHC’s Technical Moat

DeepSeek V4 can drive trillion-level parameters under limited computational power, thanks to three key innovations:

1. Engram of conditioned memorySeparating factual knowledge from dynamic reasoning, enabling the model to “accurately navigate” through million-word-long texts, achieving retrieval accuracy in the Needle-in-a-Haystack test of 97%。
2. Manifold Constraint Hyperconnection (mHC)By constraining gradient signal fluctuations within a factor of two using a mathematical framework, it solves the gradient explosion problem in trillion-parameter training, reducing training costs to approximately $10 million.
3. DSA Sparse AttentionReplaces traditional dense attention, reducing computational cost for long contexts by approximately 50%.

API Service Upgrade and “Thinking Mode”

DeepSeek API has been synchronously updated and introduces flexible calling parameters:

• Supports two modesBoth V4-Pro and Flash provide “thinking mode” and “non-thinking mode”.
• Adjustable intensityUsers can via reasoning_effort Thinking intensity setting (high/max). Officially recommended to enable in complex Agent scenarios. max Mode
• Interface Exit Reminderexisting deepseek-chat and deepseek-reasoner The interface will be July 24, 2026Support officially discontinued; developers advised to switch to as soon as possible deepseek-v4-pro or deepseek-v4-flash。

Hardware and Open Source: Turning to the Huawei Ecosystem and Apache 2.0

Against the backdrop of geopolitical factors restricting high-end GPU supply, DeepSeek V4 demonstrated Huawei Ascend 950PR and Cambricon MLU Deep optimization of chips. When running V4, the Huawei Ascend 950PR’s performance even reaches 2.87x that of the NVIDIA H20.

Model weights have been synchronized and published at Hugging Face and ModelScope, adopt Apache 2.0 The protocol is open-source, meaning developers worldwide can use it for free and customize it for their own needs or commercial purposes.

Disruptive pricing advantage

Not surprisingly, DeepSeek V4 once again showed its “price killer” colors:

• Price comparison: The API for GPT-5 or Claude 4.5 costs approximately per million input tokens NT$480 – NT$650。
• DeepSeek V4 Estimated PricingOnly about per million inputs NT$4.5 (at approximately $0.14 USD, directly reducing the barrier to accessing frontier AI to just 1% to 2% of competitors)

Conclusion

DeepSeek quoted “Not seduced by praise, not frightened by slander, proceeding along the righteous path, standing firmly in one’s integrity” in their official announcement, showcasing their commitment to technological innovation. The release of DeepSeek V4 is not merely about scaling up parameters, but a comprehensive restructuring of AI infrastructure, Agent application scenarios, and operational costs. With the release of open-source weights, the global AI landscape may face another round of dramatic upheaval.

Source: KOCPC Chinese