DeepSeek released an updated public-beta API for DeepSeek-V4-Flash alongside immediate open-weights availability, marking a significant performance leap achieved through post-training rather than architectural changes. The model retains its 284B total / 13B active parameter structure but shows substantial improvements in agentic capabilities and coding benchmarks.
- DeepSeek-V4-Flash API launched with upgraded agent capabilities surpassing the V4-Pro-Preview, supporting the Responses API format and adaptation for Codex.
- Terminal-Bench scores jumped from 56.9 to 82.7 (+25.8 points), while GDPval-AA v2 Elo increased from 1189 to 1559 without changes to model size or architecture.
- Open weights were released under the MIT license, featuring 256 routed experts with 6 active per token and a DSpark speculative decoding module.
- The release intensified price competition, positioning DeepSeek on the intelligence vs. cost Pareto frontier at $0.14/$0.28 per 1M tokens with a 98% cache-hit discount.
The update demonstrates that post-training optimizations can significantly enhance model utility for complex tasks while maintaining competitive pricing against proprietary alternatives.