DeepSeek V4.1 Flash (Scaleway)

DeepSeek V4.1 Flash is an open-weights Mixture-of-Experts language model with a 552B-parameter backbone designed to handle very long contexts and agent-style workloads efficiently. Its Causal Encoder-Decoder architecture activates roughly 8B parameters per input token during prefill and around 16B during generation, while a redesigned KV-cache scheme brings the global footprint down to about 890 bytes per token — roughly one quarter of the prior V4 Flash — through sparse attention sharing and FP4 caching. These design choices make the model well suited to long sessions, large repositories, and tool-call histories where most of the cost comes from re-reading and re-using context rather than generating new tokens.

The model is released by DeepSeek under an open license, and its stable API identifier is deepseek-flash, with legacy V4 Flash aliases kept temporarily compatible. V4.1 Flash is intended for practical agent and coding use cases that demand extended context, repeated cache reads, and tool use; benchmarks reported by DeepSeek evaluate it under maximum reasoning effort at a one-million-token context, and the architecture's input-heavy efficiency profile is its main differentiator against broader peers. For practitioners, it fits projects where long-context retrieval, reasoning, and tool orchestration dominate the workload, and where the smaller active-parameter footprint and compressed KV cache translate into faster and cheaper agent runs.

Eden AIscaleway/deepseek-v4.1-flashdeepseek-flash

Open provider page Provider docs

Quick Info

Provider: Eden AI
Model key: scaleway/deepseek-v4.1-flash
Release date: Sep 10, 2026
Last updated: Sep 10, 2026
Knowledge cutoff: 2025-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.67086
Output token cost: $2.012581

Limits

Output tokens: 384,000 tokens
Context window: 256,000 tokens

Transparent token rates

Compare DeepSeek V4.1 Flash (Scaleway) pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about DeepSeek V4.1 Flash (Scaleway)

Eden AI

Coverage

DeepSeek introduced V4.1-Flash on September 9, 2026, as the smallest model in a new architecture family with native visual understanding. It is a 552B-parameter Mixture-of-Experts model using a new Causal Encoder–Decoder architecture that activates just 8B parameters for input and 16B for output, delivering benchmark results ahead of the flagship V4-Pro. The model is live on the DeepSeek API under the model name deepseek-flash with native multimodal support. V4.1-Flash cuts KV cache needs to roughly 1/4 the HBM and 1/8 the SSD storage of the previous generation, directly reducing cache-hit costs that dominate agent workloads. DeepSeek announced new lower API pricing effective September 10, 2026, with off-peak rates at 50% of peak, and is phasing out V4-Pro by routing deepseek-v4-pro requests to V4.1-Flash. Partners WorkBuddy, CodeBuddy, and OpenCode fully support the model at launch.

Eden AI

Coverage

The DeepSeek-V4.1-Flash technical paper (arXiv 2609.19969, submitted September 17, 2026) describes a multimodal 552B-parameter MoE model supporting up to one million tokens of context. The Causal Encoder–Decoder design activates 8B parameters during prefill and 16B during decode, targeting cost efficiency for input-heavy agent workloads. Pretraining used a multimodal corpus of 45T tokens with comprehensive post-training for agentic scenarios. KV cache compression combines cross-layer reuse in Compressed Sparse Attention 2 (CSA2) with FP4 caching, reducing the global HBM-resident footprint to 890 bytes per token, about 1/4 of V4-Flash. A deployment optimization called SWA Bounded Replay cuts the persistent SSD/host-memory footprint to roughly 1/8 of the previous Flash generation. Checkpoints are publicly available on Hugging Face for self-hosting.