DeepSeek V4.1-Flash: An Encoder-Decoder and 890 Bytes of KV per Token

DeepSeek has abandoned the symmetric layer stack behind V4-Flash and released V4.1-Flash with a Causal Encoder-Decoder architecture. The model carries 552B backbone parameters but activates only 8B when reading input and 16B when generating output. The target is clear: agent workloads, where the input runs to tens of thousands of tokens and the answer is short.

Key takeaways

• 552B backbone parameters, a 94 percent jump from V4-Flash's 284B.
• 40 layers split into 20 encoder layers and 20 decoder layers — an asymmetric design.
• Global KV cache compressed to 890 bytes per token, about a quarter of its predecessor's.
• 1M-token context window, open weights under the MIT licence.
• Off-peak pricing: $0.15 per million input tokens and $0.60 per million output tokens.

Asymmetry instead of a uniform stack

A classic LLM pushes input and output through the same stack of layers. CED splits those roles: 20 layers read and process the input, the next 20 generate the answer. Hence the gap in activations — prefill^?Prefill: The phase in which the model processes the entire input in one pass, before producing the first output token. With long prompts this is where most of the compute goes. fires 8B parameters, decoding 16B. On top of the backbone sit 196B parameters in sparsely accessed Engram conditional-memory modules.

890 bytes per token

The second change concerns the attention cache. CSA2, the second generation of Compressed Sparse Attention, uses three static attention modes — Full, Reindex and Reuse — and keeps the main KV cache in FP4 (E2M1) format. The result: 890 bytes per token instead of roughly four times that in V4-Flash.

Close to the top, far below on price

On the agentic DeepSWE v1.1 benchmark the model scores 74.2 — marginally ahead of Claude Opus 5 and GPT-5.6 Sol. That is a tie within noise, but at a price tag from a different league. On reasoning it posts 90.9 on GPQA Diamond. Chinese analyses note the training recipe stayed conventional — SFT, RL and online distillation — with the gains coming mostly from the data pipeline rather than new algorithms.

Why it matters

Separating encoder from decoder is an admission that the workload has changed. Chat means short input and long output, an agent is the reverse. Optimising both with one stack of layers stops paying off. If the approach holds up, rivals will either have to copy the split or explain why they are paying for memory they do not need.

What's next?

• The weights are on Hugging Face under the MIT licence, so independent KV and throughput measurements are a matter of weeks.
• Terminal-Bench figures reported by DeepSeek and by secondary media diverge across benchmark versions — independent replication is worth waiting for.

Pulling the weights from the official repository comes down to a single call:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="deepseek-ai/DeepSeek-V4.1-Flash",
    local_dir="./DeepSeek-V4.1-Flash",
)

Sources

• Hugging Face — DeepSeek-V4.1-Flash model card
• VentureBeat — DeepSeek-V4.1-Flash debuts with $0.003/1M off-peak cached-input rate and benchmarks eclipsing GPT-5.6 Sol, Claude Opus 5
• 钛媒体 — DeepSeek「再一次推倒重来」