DeepSeek Releases V4: Open-Source Frontier Model at a Fraction of the Price
DeepSeek's V4-Pro and V4-Flash arrive with 1.6 trillion parameters, a 1M-token context window, and Apache 2.0 licensing — matching near-frontier performance at 21x less cost than Claude Opus 4.7.
DeepSeek, the Chinese AI lab that sent shockwaves through Silicon Valley in early 2025, has released its long-awaited V4 series. The two-model family — V4-Pro and V4-Flash — dropped on April 24, 2026, on Hugging Face under an Apache 2.0 license, making it the largest open-weights model currently available from any lab worldwide.
The flagship V4-Pro packs 1.6 trillion total parameters with a Mixture-of-Experts (MoE) architecture that activates only 49 billion per token, giving it frontier-class reasoning at dramatically lower inference cost. V4-Flash follows with 284 billion total parameters and 13 billion activated. Both models ship with a 1 million-token context window and support up to 384,000 tokens of output — enough to process entire codebases in a single prompt.
The architecture introduces Compressed Sparse Attention (CSA) combined with Heavily Compressed Attention (HCA), a hybrid that reduces memory overhead to just 10% of V3.2's KV cache and cuts single-token inference FLOPs to 27% of the previous generation. On coding benchmarks, V4-Pro leads the open-source field with a Codeforces rating of 3,206 — beating GPT-5.4's 3,168 — and scores 93.5 on LiveCodeBench. On SWE-bench Verified it lands at 80.6%, within 0.2 points of Anthropic's Claude Opus 4.6.
The economics are striking. V4-Pro is priced at $1.74 per million input tokens and $3.48 per million output tokens — roughly 8.6 times cheaper than GPT-5.5 and 21 times cheaper than Claude Opus 4.7. V4-Flash undercuts all comparable small models at just $0.14 input and $0.28 output, beating GPT-5.4 Nano. Models are available through the DeepSeek API and at chat.deepseek.com, in addition to the open weights on Hugging Face.
DeepSeek's release comes after three delays spanning nearly four months, and follows a turbulent period in which the original V3 release rattled Western AI stock markets. The V4 family validates the compute-efficiency strategy that has made DeepSeek disproportionately influential relative to its size, and positions open-source AI as a credible alternative to closed frontier systems for production enterprise workloads.
