DeepSeek V4 Release
DeepSeek released V4 Pro and V4 Flash, its first major architecture refresh since V3, with 1M-token context and hybrid reasoning modes.
Read MoreHighlights
- DeepSeek released V4 Pro and V4 Flash, featuring a 1M-token context and hybrid reasoning modes.
- V4 Pro is ranked as the #2 open-weights reasoning model, with strong long-context and agentic performance.
- DeepSeek V4 introduces a new long-context attention system with significant KV-cache reduction.
Models
Research
DeepSeek V4 Technical Report
DeepSeek V4's technical report is considered one of the most important model papers of the year, detailing advancements in long-context and agentic coding performance.
Read MoreTools
DeepSeek V4 Quantization
DeepSeek V4 uses mixed FP4 + FP8 quantization, allowing the full model to fit on a single 8×B200 node.
Read MoreProducts
DeepSeek V4 Pricing
DeepSeek V4 Pro and Flash pricing announced, with potential for price reduction once Huawei Ascend 950 supernodes are deployed.
Read More