DeepSeek V4 preview narrows the gap with open-weight MoE models
V4-Pro at 1.6T params (49B active) and V4-Flash at 284B, both 1M-context, released open-weight on Hugging Face
Summary
Deepseek released a preview of V4 on 24 April 2026, listing deepseek-v4-pro and deepseek-v4-flash as API model IDs and publishing open weights on Hugging Face. V4-Pro is a 1.6-trillion-parameter mixture-of-experts model with 49B activated; V4-Flash is 284B with 13B activated; both carry 1M-token context. DeepSeek claims V4 closes the gap with frontier models, beating some GPT-5.2 and Gemini 3.0 Pro results on reasoning. It remains a preview with no finalisation date. The open-weight release, far cheaper per token, sharpens the open-vs-closed split and runs against US chip-export constraints DeepSeek trains under.
By the numbers
- 1.6T, V4-Pro parameters (49B activated).
- 284B, V4-Flash parameters (13B activated).
- 1M tokens, context window, both models.
- 24 Apr 2026, preview release.
- Open weights, published on Hugging Face.
Why it matters
A frontier-adjacent open-weight model from a Chinese lab undercuts the closed labs on price and removes the access gate the US is building around top US models, the asymmetry the Anthropic export-control order exposed. It keeps China competitive at the frontier despite chip controls.
What to watch
- Independent benchmarks vs. GPT-5.5, Gemini 3.5 and Opus 4.8.
- A finalised (non-preview) V4 and pricing.
- Whether US export policy targets open-weight Chinese models.