FrontierSeptember 10, 2026via The Decoder
New Deepseek model V4.1-Flash cuts memory needs for AI agents
Why it matters
DeepSeek V4.1-Flash is a frontier model release that combines extreme parameter efficiency (16B active per token from 552B total), dramatic memory optimization (KV cache cut by 75%), and competitive capability on hard benchmarks — all under MIT license. For practitioners, this reshapes the cost equation for agentic AI; for enthusiasts, it's the latest salvo in the lab race's efficiency arms race.
Key signals
- 552 billion total parameters; 16 billion active per token (MoE)
- KV cache memory reduced to 25% of predecessor
- DeepSWE benchmark: narrowly beats Opus 5 and GPT-5.6 Sol
- Multimodal capability
- MIT license (open-weight distribution)
- Positioning: cheaper AI agents
The hook
DeepSeek just cut KV cache memory to 25% while matching Opus 5 on code. That changes the economics of every agent deployment.
Deepseek releases V4.1-Flash, a multimodal model with 552 billion parameters that cuts KV cache memory to a quarter of its predecessor. On the DeepSWE coding benchmark, it narrowly beats Opus 5 and GPT-5.6 Sol, even though only 16 billion parameters are active per token. The model ships under the MI…