DeepSeek Launches V4-Pro: Open-Source Model Outperforms Claude Opus 4.6 and GPT-5.4

DeepSeek's New Flagship Model Goes Live

On April 24, 2026, Chinese AI startup DeepSeek unveiled a preview of its V4 model family. The flagship DeepSeek-V4-Pro outperformed Claude Opus 4.6 and GPT-5.4 across several benchmarks, establishing itself as the strongest open-source model currently available.

"🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Flash: 284B total / 13B active params." — DeepSeek (@deepseek_ai), original post

Why This Matters

An open-source model competing head-to-head with the best proprietary offerings from Anthropic, OpenAI, and Google represents a significant shift in the AI landscape. With V4-Pro's weights and architecture publicly available, researchers and companies worldwide can deploy and fine-tune a top-tier system without relying on proprietary APIs. For the crypto industry — where AI agents are increasingly being integrated into trading strategies and DeFi protocols — this provides a powerful and cost-efficient tool.

Architecture and Scale

V4-Pro contains approximately 1.6 trillion parameters, but only 49 billion are activated during each inference step. The smaller sibling, V4-Flash, has 284 billion total parameters with 13 billion active per step.

Both models use a Mixture of Experts (MoE) architecture, where only the neural sub-networks relevant to the current task are engaged when processing each token. This approach significantly reduces computational costs while maintaining performance comparable to fully dense architectures.

Pre-training was conducted on a corpus exceeding 32 trillion tokens. The team then applied staged fine-tuning, dedicating separate blocks to coding, mathematics, logical reasoning, and instruction-following. The final model unifies these capabilities through distillation.

Long Context at a Fraction of the Cost

The defining feature of V4 is its dramatic reduction in the cost of processing long sequences. While a 1 million token context window is available from competitors, using it typically involves substantial costs and latency.

According to DeepSeek, V4-Pro requires only about 27% of the compute and 10% of KV-cache memory compared to V3.2 when working at maximum context length. V4-Flash achieves even greater efficiency — approximately 10% of compute and 7% of memory.

These gains stem from a hybrid attention architecture employing two data compression mechanisms to reduce overhead when processing long texts. The team also utilized specialized hyperconnections for training stability and the Muon optimizer for faster convergence.

Three Reasoning Modes and Agent Capabilities

The V4 models support three distinct reasoning modes:

• Non-think — instant responses to straightforward queries without additional analysis;
• Think High — deep analysis for complex tasks and planning;
• Think Max — full mode where the model details every step and evaluates all possible paths.

For agentic tasks, Think Max now preserves the chain of intermediate steps within a single task. In the previous version, portions of this context were lost during user interactions.

Benchmark Results

According to DeepSeek's published data, V4-Pro delivered competitive performance across a broad range of evaluations:

• Coding: a Codeforces rating of 3206 — 23rd among human programmers worldwide, on par with GPT-5.4;
• Mathematics: 95.2 on HMMT 2026 and 89.8 on IMOAnswerBench, ahead of most competitors;
• Knowledge (SimpleQA Verified): 57.9 (Opus 4.6 scored 46.2, though Gemini 3.1 Pro reached 75.6);
• Software development (DeepSeek internal test): 67% — between Sonnet 4.5 (47%) and Opus 4.5 (70%);
• Agentic scenarios: V4-Pro-Max achieved 80.6% on SWE Verified and 67.9% on Terminal Bench.

The V4 models were specifically trained on practical work scenarios including data analysis, report generation, document editing, and iterative web search with tool use. In an internal survey of 85 DeepSeek developers and researchers, 52% said they were ready to adopt V4-Pro as their primary coding model, while an additional 39% indicated they were leaning toward doing so.

The DeepSeek V4 release came just one day after OpenAI launched GPT-5.5 on April 23, which was positioned as "a new level of intelligence for real work and agent management." The race among leading AI labs continues to intensify.