DeepSeek Launches V4.1-Flash, Quickly Picked Up by Multiple Inference Providers

On Sept 18, DeepSeek officially released DeepSeek-V4.1-Flash, the smallest model in its new architecture family: a multimodal MoE built for long-horizon agent workloads, paving the way for larger family members. Weights are open on Hugging Face, and several third-party inference providers launched it the same day.

Confirmed

Architecture: 552B MoE with an asymmetric causal encoder–decoder, activating only 8B parameters on input and 16B on output, with native vision understanding (m1, m4, m5)
Long context and efficiency: up to 1M-token context, KV cache compressed to 890 bytes/token — a quarter of the previous footprint (m1, m5)
Openness: weights available on Hugging Face (m1)
Third-party rollouts: Nebius Token Factory launched it for long-input, high-throughput workflows (m4); Runware launched it and disclosed input token pricing (m3); Inco reported 532 tokens/s output, ranking first on its platform per Artificial Analysis (m2)

Why it matters

The asymmetric activation (8B in / 16B out) balances long-input cost with generation quality, and the heavily compressed KV cache lowers the barrier for million-token context applications.
As the smallest member of the new family, it validates the architecture and deployment path for larger models; same-day adoption and a speed-chart top spot show strong ecosystem uptake.

2026-09-18 ~ 2026-09-19 · 5 related posts

Full story(2 episodes)→

Episode 1: DeepSeek Flash Becomes Default Workhorse, Cutting AI Costs 90%(2026-09-18, 2 posts)
Episode 2: DeepSeek Launches V4.1-Flash, Quickly Picked Up by Multiple Inference Providers(2026-09-18, 5 posts)

Primary sources

DeepSeek releases V4.1-Flash: 552B MoE with 1M context, KV cache cut to 890 bytes/token — deepseek-ai · 2026-09-18 11:17 (UTC+08:00)
DeepSeek-V4.1-Flash lands on Nebius: 552B MoE activating 8B with 1M context — Arindam_1729 · 2026-09-18 03:10 (UTC+08:00)
DeepSeek V4.1 Flash hits 532 tokens/s on Inco, fastest output on Artificial Analysis — songhan_mit · 2026-09-18 05:36 (UTC+08:00)

[source] DeepSeek-V4.1-Flash lands on Nebius: 552B MoE activating 8B with 1M context — Arindam_1729 · 2026-09-18
DeepSeek releases V4.1-Flash: smallest model in new family with native vision, $0.15/M tokens — aziz4ai · 2026-09-18
[source] DeepSeek V4.1 Flash hits 532 tokens/s on Inco, fastest output on Artificial Analysis — songhan_mit · 2026-09-18
[source] DeepSeek releases V4.1-Flash: 552B MoE with 1M context, KV cache cut to 890 bytes/token — deepseek-ai · 2026-09-18
DeepSeek ships 552B MoE DeepSeek-V4.1-Flash with 1/4 KV cache, phasing out V4-Pro — dl_weekly · 2026-09-19