DeepSeek V4 goes open-source with million-token context, undercutting US rivals on price
Chinese AI lab DeepSeek has released V4-Flash and V4-Pro, open-source models under MIT license with a native 1M-token context window. The hybrid attention architecture cuts inference costs to a fraction of US competitors, per Fello AI. European developers can now deploy frontier-level AI at €0.20 ($0.22) per million input tokens for V4-Flash.
Bottom line — DeepSeek V4's open-source release gives European AI teams a 36x cheaper alternative to GPT-5.4 for long-context tasks.
Go deeper (8)
- V4-Flash has 284B total parameters (13B active) and V4-Pro has 1.6T total (49B active), per the technical report.
- The hybrid attention system uses Compressed Sparse Attention and Heavily Compressed Attention, reducing KV cache by 90% at 1M context compared to V3.2, according to Fello AI.
- On Terminal Bench 2.1, V4-Flash-0731 scores 82.7, beating V4-Pro preview's 72.1, per DeepSeek's own model card.
- V4-Pro-Max achieves 93.5% on LiveCodeBench and a Codeforces rating of 3,206 (23rd among humans), per the tech report.
- Pricing for V4-Flash is €0.20 ($0.22) per million input tokens off-peak, with cache hits at €0.006 ($0.007), per DeepSeek's API docs.
- The US National Defense Authorization Act for 2026 bans DeepSeek from Department of Defense devices, Wikipedia reports. Australia also banned it from government systems.
- DeepSeek also released DeepSeek Harness, an open-source agent harness with a plugin architecture, per The Register.
- Huawei announced its Ascend 950 AI chips will fully support DeepSeek V4, enabling deployment on Chinese domestic hardware, per Fello AI.