DeepSeek ships V4.1-Flash with a novel prefill/decode-split MoE architecture, retiring old Flash naming
2026-09-10
DeepSeek released DeepSeek-V4.1-Flash, described by Latent Space as a 763B-total-parameter MoE with an 8B-active prefill / 16B-active decode split (a causal encoder–decoder-style design) plus vision support and Sliding-Window Attention Bounded Replay, cutting KV-cache footprint to roughly 1/8 of V4 Flash. DeepSeek's API docs confirm the legacy `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` names are retired and now served by V4.1-Flash at the Flash price, while V4 Pro API access continues past its planned September 14 cutoff by user demand. DeepSeek also opened a developer preview of an agent-tool integration product called 'DeepSeek Harness.'
Significance 3: A genuine open-weights architecture innovation (prefill/decode parameter split, large KV-cache reduction) that a respected curator calls significant, but per the curator's own account it trails other open models on published benchmark headlines and no independent numbers accompany the release — worth reading, not yet landscape-changing.
Implications · machine-drafted, not owner judgment
If the KV-cache reduction holds up under independent testing, V4.1-Flash could meaningfully cut serving cost for long-running agentic workloads — directly relevant to the bulk-cheap verdict, which currently rests on deepseek-v4 pricing and speed. It also keeps DeepSeek's research cadence visible after a year of GLM and Kimi taking the open-weights spotlight. Nothing here yet touches the local-hardware question — the model is still far too large for a 64GB machine.
- Independent Artificial Analysis pricing/speed numbers for V4.1-Flash
- A published SWE-bench Verified or Terminal-Bench score against Opus 5 / GPT-5.6 / GLM-5.3
Sources
- curator [AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale retrieved 2026-09-12
- primary DeepSeek API docs — model naming and V4 Pro continuation notice retrieved 2026-09-12
Models
Affected axes
Coding · Agentic / tool use · Long context · Cost