A year ago, DeepSeek V3 caught the AI industry off guard - matching GPT-4 class performance at a fraction of the usual compute cost, while running as a local model on developer machines within days of release. Now the Chinese lab has released a preview of V4, its next-generation open-source model, claiming it can compete with closed-source frontier systems from Anthropic, Google, and OpenAI.
The preview emphasis is on coding. That's not accidental - coding benchmarks have become the standard way to evaluate whether a model can handle structured, precise, verifiable tasks. DeepSeek says V4 marks a major leap over V3 specifically in this area, putting it in direct competition with the developer-focused work happening at ChatGPT and Claude.
No independent benchmark numbers have been published yet. The "competitive with US rivals" framing is self-reported for now, which is standard practice for a preview - full evaluations typically ship alongside the official release. Given DeepSeek's history, that claim deserves serious attention rather than reflexive skepticism.
What Open-Source Actually Means Here
The open-source angle is what gives this weight beyond a typical model announcement. Open-source in this context means the weights - the actual numerical parameters that make the model work - get published for anyone to download and run. Developers can deploy V4 on their own servers, fine-tune it (train it further on their own specialized data), and build products without paying per-query API fees. That changes the cost math significantly for any company building a coding tool, code assistant, or developer product on top of AI.
Per the Verge report, the official V4 release and full benchmark disclosures are expected to follow the preview. Right now the lab is signaling it's back in the frontier conversation - the numbers come later.
Pressure on Western Pricing
If V4 matches current top models on coding tasks and ships openly, the argument for paying premium API rates weakens. Anthropic and OpenAI have both leaned hard into developer use cases over the past six months. A strong open-source alternative doesn't just compete on capability - it gives enterprise buyers a credible negotiating position on pricing. For teams building anything that relies heavily on AI-generated code, V4 is worth tracking closely when the full release drops.