DeepSeek previews new AI model that 'closes the gap' with frontier models

· Source ↗

TLDR

  • DeepSeek released V4 Flash and V4 Pro previews, its largest open-weight models yet, at prices undercutting major frontier competitors.

Key Facts

  • V4 Pro has 1.6 trillion total parameters (49B active), making it the largest open-weight model available; V4 Flash has 284B (13B active).
  • Both models match GPT-5.4 on coding benchmarks but trail GPT-5.4 and Gemini 3.1 Pro on knowledge tests by an estimated 3-6 months.
  • V4 Flash costs $0.14/M input tokens; V4 Pro costs $0.145/M input and $3.48/M output, undercutting Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro.
  • Both models are text-only with 1 million token context windows; multimodal support is absent unlike many closed-source peers.

Why It Matters

  • The V4 series is DeepSeek’s most capable open-weight release, narrowing the performance gap with top closed models while maintaining its price advantage.
  • The launch follows U.S. accusations that China stole American AI labs’ IP, and prior claims by Anthropic and OpenAI that DeepSeek distilled their models.

Ram Iyer, TechCrunch · 2026-04-24 · Read the original