Hasty Briefsbeta

Bilingual

DeepSeek V4: The Open-Source Model Frontier Labs Feared

9 hours ago
  • #coding benchmarks
  • #open-source AI
  • #cost efficiency
  • DeepSeek V4 is an open-source model available under an MIT license with minimal commercial restrictions.
  • Priced at $0.30 per million output tokens, it is significantly cheaper than competitors like Claude Opus 4.7 and GPT-5.5, by a factor of 83 to 100 times.
  • Achieves high performance in coding benchmarks: 80.6% on SWE-bench Verified, 93.5 Pass@1 on LiveCodeBench, and a Codeforces rating of 3206.
  • Utilizes a 1.6-trillion-parameter Mixture of Experts architecture, activating only 49 billion parameters per token, which improves inference efficiency.
  • Reduces inference FLOPs to 27% of its predecessor and cuts KV cache occupancy to 10% at 1 million token context.
  • Challenges closed-model dominance in agentic coding by offering competitive quality at a fraction of the cost.
  • Has caveats including less transparent benchmarking, Chinese origin with data governance concerns, and self-hosting complexities.
  • Resets the price floor for advanced coding intelligence, potentially forcing competitors to lower prices or enhance capabilities.