DeepSeek-V4.1-Flash

DeepSeek introduced DeepSeek-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 using a Causal Encoder-Decoder design that activates 8B parameters for input and 16B for output, with new pretraining and larger-scale RL post-training delivering benchmarks ahead of DeepSeek-V4-Pro. The model went live on the DeepSeek API as deepseek-flash with native multimodal support. The release brings major KV cache compression: V4.1-Flash needs only 1/4 the HBM and 1/8 the SSD storage of the previous generation, cutting cache-hit costs that dominate agent workloads. V4-Flash and V4-Flash-Vision-Exp were retired, with their model names temporarily routed to V4.1-Flash; V4-Pro was being phased out (later continued). New peak/off-peak pricing took effect September 10, 2026, with off-peak at 50% of peak rates, and the weights are available on Hugging Face.

DeepSeek's API documentation mirrors the V4.1-Flash launch announcement, describing it as a 552B-parameter MoE with a Causal Encoder-Decoder architecture activating 8B parameters on input and 16B on output. New pretraining and larger-scale RL post-training are credited with benchmark results ahead of flagship V4-Pro, and the model ships with native multimodal visual understanding. The API entry point is set to deepseek-flash. The page confirms API-side behavior: V4-Flash and V4-Flash-Vision-Exp are retired with deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily routed to V4.1-Flash, and V4-Pro requests are scheduled to route to V4.1-Flash at V4.1-Flash rates from September 14, 2026. KV cache footprint drops to 1/4 the HBM and 1/8 the SSD of the prior generation, with new off-peak pricing at 50% of peak effective September 10, 2026.

Flowtivity's independent analysis (September 10, 2026) reports that DeepSeek-V4.1-Flash beats GPT-5.6 Sol and Claude Opus-5.0 on four of the hardest agentic benchmarks: DeepSWE v1.1 (74.2 vs 73.0), AutomationBench (54.8 vs 45.8), Agents' Last Exam (31.8 vs 26.7), and CyberGym (88.1 vs 84.5). The 552B-parameter MoE activates 8B per token during prefill, carries a 1M-token context, reads images natively, and costs $0.30 per million input tokens at peak. The piece frames the release as a cost-efficiency play: V4-Pro was being retired on September 14, 2026 in favor of V4.1-Flash, with DeepSeek stating V4.1-Flash surpasses V4-Pro across performance, cost, speed, and task completion time. The model ships under MIT license with 384K maximum output tokens, and V4.1 Pro is positioned as an upcoming successor.

Vals AI lists DeepSeek V4.1 Flash as an open-weight model released September 10, 2026, with a 1M-token context window, 384K max output tokens, and $0.30/1.20 per million input/output token pricing. It ranks 24 of 45 on the Vals Index at 51.32% accuracy, with its best result being 1 of 35 on SkillsBench. Native image input is supported, while video and file inputs are not. Per-category benchmark standings include Code Migration 45.62% (15/75), CyberBench v1.1 73.69% (9/45), EMB 57.21% (37/72), Finance Agent v2 53.48% (30/76), Legal Research Bench 41.35% (22/75), MedCode 41.17% (56/106), MedScribe 85.50% (26/108), ProofBench v1.1 54.00% (26/48), and SAGE 47.8%, giving a cross-vendor independent view of the model's capabilities.

The DeepSeek API change log entry for September 10, 2026 officially announces the V4.1-Flash release as the smallest model in the new architecture family with native multimodal visual understanding. Published benchmark numbers include GPQA Diamond 90.9, HLE 36.8, Codeforces rating 3471, MathArena Apex 65.6, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, CyberGym 88.1, and Agents' Last Exam 31.8. API prices were reduced with the release. The log notes V4-Flash and V4-Flash-Vision-Exp are retired with their model names temporarily routed to V4.1-Flash via deepseek-flash, and reverses the earlier V4-Pro phase-out by continuing V4-Pro API service after September 14, 2026 with unchanged billing. The entry also includes multimodal and agentic benchmark results such as Chartography 78.9, BabyVision 89.6, and HLE with tools 63.9.

arXiv preprint 2609.19969, submitted September 17, 2026 by the DeepSeek-AI team, is titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression." It serves as the primary technical reference for the model, authored by a large DeepSeek-AI contributor list covering the CED architecture, SWA Bounded Replay, CSA2 attention, and FP4 KV caching techniques. The paper documents the same headline numbers echoed elsewhere: a 552B-parameter MoE with 8B active parameters on input and 16B on output, 1M-token context, and KV cache compressed to roughly 1/4 the HBM and 1/8 the SSD of the prior Flash generation. Submitted one week after the September 10, 2026 release, it is the formal technical record behind the launch claims.

The official Hugging Face model card details DeepSeek-V4.1-Flash as a multimodal MoE with 552B backbone parameters and support for up to one million-token contexts, processing images and text natively. Architecture is a 40-layer Causal Encoder-Decoder (20-layer causal encoder plus 20-layer decoder) that activates 8B parameters per token during prefill and 16B during decode, targeting cost efficiency for input-heavy agentic workloads. Key innovations include SWA Bounded Replay, Compressed Sparse Attention 2 (Full/Reindex/Reuse modes with a Hierarchical Sparse Indexer), and FP4 main KV caching that brings the global footprint to roughly 890 bytes per token, about 1/4 of DeepSeek-V4-Flash. Each MoE layer uses 1 shared plus 384 routed experts with 6 routed active per token, plus Engram conditional memory (196B) and DSpark speculative decoding.