DeepSeek cuts API pricing 60% and open-sources its V4 base weights
The Chinese lab continues its cost-disruption strategy: V4 base model weights land on Hugging Face under a permissive license while hosted inference gets dramatically cheaper. Demo content.
DeepSeek has published the base weights of DeepSeek-V4 on Hugging Face under a permissive license and simultaneously cut hosted API prices by 60%, extending the cost-disruption playbook that has repeatedly forced Western labs to respond.
The numbers
- ▸Input tokens now priced at a fraction of comparable frontier APIs
- ▸V4 base weights: 671B parameters, MoE architecture, 42B active
- ▸Quantized community builds already running on dual consumer GPUs
License details
The release uses a permissive license with no revenue-based restrictions, making V4 immediately attractive for startups that resisted more encumbered open-weight options.
# Community 4-bit build, single command via llama.cpp
llama-server -m deepseek-v4-q4_k_m.gguf --ctx-size 65536Analysts expect renewed pricing pressure across the API market within weeks, particularly on high-volume extraction and summarization workloads where frontier-level reasoning is unnecessary.
Source: DeepSeek
