An open-source model that competes with GPT-5.5 and Claude Opus 4.7 in coding, reasoning, and math benchmarks — but costs 7x less. This is not marketing hype. DeepSeek V4 was released on April 24, 2026 under the MIT license, with full code available on GitHub.
Two Variants, Two Different Goals
V4-Pro: The flagship model with 1.6 trillion total parameters, but only 49 billion active at any given time. It uses a Mixture of Experts (MoE) architecture — think of it as a large hospital with a thousand specialists, but only 2-3 relevant experts are called for each patient. High quality without consuming all resources.
V4-Flash: A lighter version with 284 billion total parameters and just 13 billion active. Designed for when speed and cost matter more than peak accuracy. Great for chatbots, summarization, and everyday tasks.
Both support a 1 million token context window — roughly 14 books worth of text in a single input.
Pricing That Challenges the Industry
V4-Pro input: $0.145 per million tokens. The same task with GPT-5.5 or Claude Opus 4.7 costs roughly 7x more.
V4-Pro output: $1.74 per million tokens — approximately 6x cheaper than Western competitors.
Benchmark Performance
V4-Pro scored 80.6% on SWE-Bench, a software engineering benchmark that gives the model real bugs from GitHub repositories and asks it to fix them. That means the model correctly solved more than 4 out of every 5 bugs.
On math reasoning, coding, and text comprehension benchmarks, V4-Pro performs roughly on par with closed-source frontier models. But with one fundamental difference: V4 code is completely open and free.
MIT License — Truly Open Source
DeepSeek V4 ships under the MIT license — the simplest and most permissive open-source license. You can use it commercially, modify it, distribute it, and even fine-tune and sell the model with zero restrictions.
For companies concerned about vendor lock-in, this is a real opportunity. You can self-host the model and keep all data on your own infrastructure.
Four Chinese Models in 12 Days
V4 was not alone. Within a 12-day window, four Chinese companies released frontier models: DeepSeek V4, Z.ai GLM-5.1, MiniMax M2.7, and Moonshot Kimi K2.6 — all achieving roughly the same capability ceiling at meaningfully lower costs.
But V4 stands apart for its transparency. DeepSeek publishes its training methods and technical papers, allowing the research community to learn and build better models.
Why This Matters
A few years ago, the common belief was that building frontier models required billions of dollars, thousands of GPUs, and massive teams. Only OpenAI, Google, Anthropic, and Meta could do it.
DeepSeek challenged that belief — not once, but repeatedly. When a company with fewer resources can build a model that matches the giants at 7x lower cost, the question changes. It is no longer “Can open source compete with closed source?” The answer is yes. The question is now: what added value do closed-source models offer to justify their higher price?
What It Means for Developers
Lower costs: If you use APIs, V4-Pro can dramatically reduce your expenses.
Independence: With the MIT license, you can self-host. Critical for projects where data privacy matters.
Massive context: 1 million tokens means you can feed an entire medium-sized codebase to the model for bug finding or feature development.
The ultimate winners of this competition are us — the users and developers. Better, cheaper, and freer models.