DeepSeek unveils v4.1-Flash with novel encoder–decoder architecture and native vision support

DeepSeek’s latest release, v4.1-Flash, introduces a novel causal encoder–decoder architecture that departs from the Transformer lineage, alongside native vision support. The company frames this as a deliberate step toward more efficient and creative use of context, rather than chasing incremental benchmark gains.

The model’s parameter budget is reported as 763B total, with an 8B active prefill module handling input tokens and a 16B active decode module handling output tokens. This prefill/decode separation yields a sparsity of roughly 1–2%, according to the technical report. DeepSeek highlights accompanying innovations such as Sliding-Window Attention with Bounded Replay, which together are said to reduce the KV cache footprint to as little as one-eighth that of V4 Flash, improving throughput and lowering inference costs for long-running agentic workflows.

Independent evaluations cited by AINews indicate that v4.1-Flash surpasses DeepSeek V4 Pro 0813 on the Artificial Analysis Intelligence Index (score: 40), placing it just below GLM-5.3-Flash and above the latest V4 Pro. Artificial Analysis also reports pricing at $0.30 per 1M input tokens and $1.20 per 1M output tokens, with cached input at $0.006 per 1M tokens and an additional 50% off-peak discount.

On the Vals Index, v4.1-Flash is described as the new top open-weight model, ahead of Kimi K3, at a cost of $0.30 per test. The Vals evaluation used a 1M-token context window, 384 max output tokens, temperature 1, default top-p/top-k sampling, and high reasoning effort. Baseten announced day-zero support and summarized the positioning as “smarter, faster, and more efficient” than DeepSeek V4 Pro 0813, with text and vision inputs, US-only availability, and a 1M-token context window. Ollama began rolling out the model to Max and Team accounts, later expanding to Pro plan subscribers.

The release notes emphasize post-training efficiency, with DeepSeek stating that improving data quality yields higher returns than novel post-training algorithms at this stage. This stance aligns with commentary from Prof. Jie Tang and Jasper Lu, who argue that data quality ROI now exceeds algorithmic novelty in post-training.