DeepSeek releases V4.1-Flash, a multimodal MoE with a much smaller KV cache, under MIT

DeepSeek published DeepSeek-V4.1-Flash. The card describes a multimodal mixture-of-experts model with 552B backbone parameters, 1M-token context, and native image and text input.

The card spends most of its length on efficiency. A 40-layer causal encoder-decoder design means the model activates 8B parameters per token during prefill and 16B during decode. Compressed Sparse Attention 2 and FP4 caching bring the global KV cache to 890 bytes per token, which the card says is roughly a quarter of DeepSeek-V4-Flash. It was pre-trained on 45T tokens. A reasoning-effort setting takes an integer from 1 to 100.

The licence is MIT. The efficiency figures are the maker's own and the card links a technical report.

The card gives no release date; the date used is the Hugging Face repo creation date.

The interesting number is bytes per token, not the leaderboard row. It is rare for a model card to lead with its memory bill.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 4: MIT licence and ungated weights; the size still needs serious hardware.
• Useful 3.5: Cheap long-context inference is the pitch, with an adjustable reasoning-effort setting.
• Notable 4: Over 4,600 likes; the V4 line is a known open reference.
• Novel 4: Causal encoder-decoder design and heavy KV-cache compression are the card's focus.