DeepSeek has released V4.1-Flash, a new open-weight flagship model featuring a novel causal Encoder-Decoder architecture. This model emphasizes extreme inference efficiency and cost-effectiveness, with a 763 billion parameter count and an 8B/16B parameter split for input and output processing. Despite not topping all benchmarks, its innovative context utilization and efficient KV cache make it particularly suitable for long-running AI agents, marking DeepSeek's return to publishing state-of-the-art research. AI
IMPACT This release highlights advancements in efficient context utilization and cost-effective inference, crucial for developing more capable and accessible AI agents.
RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- 763B-P8B-D16B
- Artificial Analysis Intelligence Index
- DeepSeek
- DeepSeek V4 Pro
- Encoder dashDecoder
- MIT license
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →