DeepSeek Harness, 39 days later

In August we audited DeepSeek Harness and listed five things to watch. Five weeks and 66,000 stars later we re-ran the same queries. One prediction landed, and only halfway.

DeepSeek Harness: The Open-Source Agent Layer Nobody Owns Yet

DeepSeek shipped the middle layer of the agent stack under MIT and told everyone to take it apart. Here is what dsh actually is, what the GitHub and npm APIs really say about its adoption, and why a free harness arrived three days before V4 output tokens got 4.5 times more expensive at peak.

DeepSeek-V4-Flash

DeepSeek-V4-Flash 0731 is official: MIT weights, 284B MoE (13B active), 1M context, and agent benchmarks above V4-Pro-Preview at $0.14/$0.28 per 1M tokens.

DeepSeek-V4

DeepSeek-V4 Preview ships two MIT-licensed MoE variants with native 1M context: V4-Pro at 1.6T params and V4-Flash at 284B, both built for agents.

DeepSeek V3-0324

DeepSeek V3-0324, launched on March 24, 2025, is a 671 billion-parameter AI model from DeepSeek, notable for its Mixture-of-Experts architecture that activates only 37 billion parameters per token, enhancing efficiency. Competing with models like GPT-4o and Claude 3.5 Sonnet, it excels in reasoning, code generation, and multilingual tasks. Despite its size, it can run on high-end consumer PCs using optimizations like 4-bit quantization, requiring components like NVIDIA RTX 4090 GPUs, 64-128 GB RAM, and fast NVMe SSDs. Although running this model on consumer hardware involves trade-offs in speed and complexity, it remains accessible thanks to its open-source MIT license, offering a democratizing force in AI development. Future updates may enhance efficiency for consumer setups, leveraging community insights and potential model refinements.