DeepSeek V4 Preview Closes Gap With Frontier Models at a Fraction of the Price

Chinese AI lab DeepSeek has unveiled two preview versions of its long-anticipated V4 model family, claiming the mixture-of-experts systems have "almost closed the gap" with frontier closed-source models from OpenAI, Anthropic, and Google. The release on April 24 makes V4-Pro, with 1.6 trillion total parameters and 49 billion active per token, the largest open-weight model publicly available.

The lineup includes V4-Flash, a leaner 284-billion-parameter model with 13 billion active, and V4-Pro for harder workloads. Both ship with one-million-token context windows and support dual reasoning modes that can be toggled between fast non-thinking responses and longer chain-of-thought outputs. DeepSeek says the models lead all open systems on math, STEM, and coding benchmarks, and trail only Gemini 3.1 Pro on world knowledge tests.

The headline architectural change is a hybrid attention mechanism that combines compressed sparse attention with heavily compressed attention to slash long-context costs. Internal numbers show V4-Pro using just 27% of the per-token FLOPs and 10% of the KV cache size of its V3.2 predecessor at the one-million-token mark, with V4-Flash dropping to roughly 10% of single-token compute. That efficiency feeds directly into pricing — V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while V4-Pro is $1.74 input and $3.48 output, undercutting GPT-5.4 and Claude offerings by an order of magnitude.

Released under an MIT license with weights on Hugging Face, the V4 family also exposes OpenAI ChatCompletions and Anthropic-compatible API endpoints, easing migration for teams already using those SDKs. Independent reviewers including Simon Willison have called V4-Pro "almost on the frontier, a fraction of the price," and analysts estimate the gap to top closed-source rivals has narrowed to roughly three to six months. With the company explicitly framing this as a preview rather than a final release, the next refinement is expected to push the open-source frontier closer still.