DeepSeek released DeepSeek-V4.1-Flash, a multimodal mixture-of-experts model. It supports contexts up to one million tokens. The model uses a causal encoder-decoder architecture. It activates 16B parameters per token during decode and 8B during prefill. It combines cross-layer KV cache reuse with FP4 KV caching. Global KV cache footprint is 890 bytes per token. Persistent KV cache footprint is one-eighth of DeepSeek-V4-Flash. The model outperforms DeepSeek-V4-Flash on agentic benchmarks. It was pretrained on 45T tokens. Checkpoints are available on Hugging Face.
What changed
DeepSeek-V4-Flash used 4x the global KV cache and 8x the persistent KV cache per token.
What it unlocks
Running long-horizon agents with one million token contexts at lower memory and bandwidth cost.
- 552B backbone parameters
- 1M token context
- 890 bytes per token global KV cache
- 45T token pretraining corpus
What you need to act on it
- access to Hugging Face model checkpoint
Sources