DeepSeek posts V4.1-Flash: 552B MoE, 8B active, 1M context

DeepSeek posted a new model card on Hugging Face for DeepSeek-V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts release under an MIT license. The headline number is not the 552B. It is 8.

The card describes 'a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder...allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.' Each MoE layer holds 1 shared expert and 384 routed experts, with 6 routed experts activated per token.

Context runs to 1 million tokens. The KV cache footprint is listed at 890 bytes per token. The card claims 'approximately 4-fold and 437-fold reductions' in KV cache relative to DeepSeek-V4-Flash and DeepSeek-V1 respectively, the larger figure carrying the compounding effect of several architectural passes across the line.

Benchmarks published on the card, run at maximum reasoning effort on the instruct model, include Terminal-Bench 2.1 at 90.6 Pass@1, DeepSWE v1.1 at 74.2 resolved, GSM8K at 93.0 EM, MMLU-Pro at 74.1, and a Codeforces rating of 3471. Pre-training used a 45T multimodal-token corpus, with context extension to 1M happening at the 34T-token mark. Four researchers we track shared the card within its first day online.