DeepSeek V4.1 Flash (Nebius)

Coursiv's comparison confirms that DeepSeek V4.1 Flash is a generational replacement for V4 Flash rather than a point update: the backbone nearly doubles to 552B parameters while active parameters per token drop from 13B to 8B on input and 16B on output, the architecture shifts from a Mixture-of-Experts decoder to a Ca The article documents DeepSeek's September 10, 2026 release and the resulting pricing change: from 04:00 UTC on September 14, 2026, all requests to the deepseek-v4-pro endpoint are routed to V4.1 Flash and billed at Flash rates until a V4.1 Pro ships, yielding roughly a 77 percent drop in cache-miss input cost and abou

The Next Web reports that DeepSeek released V4.1-Flash on September 10, 2026, publishing the weights on Hugging Face under the MIT license and announcing in an X thread that the model is positioned as the smallest member of a new causal encoder-decoder architecture family. It uses 552B total parameters with 8B active p The article documents the simultaneous retirement of V4-Pro: from 04:00 UTC on September 14, 2026, all deepseek-v4-pro requests are routed to V4.1-Flash and billed at Flash rates until a V4.1-Pro is released. DeepSeek's published benchmark table shows V4.1-Flash scoring 74.2 on DeepSWE v1.1 (ahead of Claude Opus 5's 74

BuildFastWithAI's review confirms DeepSeek V4.1 Flash is the newest Flash-series model with native multimodal visual understanding integrated into the architecture, a 1M-token context window, and up to 384K maximum output tokens, targeting coding, software agents, long-context reasoning, and high-throughput API workloa The review documents DeepSeek's official Flash pricing effective September 10, 2026: $0.003 per million cache-hit input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens off-peak, with peak pricing at double those rates. DeepSeek also confirms that V4-Pro traffic will be routed to V4.

A research-grade technical analysis published on September 10, 2026, documents that DeepSeek V4.1 Flash introduces a Causal Encoder-Decoder (CED) architecture, replacing the pure decoder MoE design of V4 Flash with 20 encoder plus 20 decoder layers. The model uses a 552B-parameter backbone with asymmetric activation of The deep dive details that V4.1 Flash cuts global KV cache requirements to roughly 890 bytes per token—about a quarter of V4 Flash's footprint—via Compressed Sparse Attention 2 modes and FP4 (E2M1) cache compression, while supporting 1M-token context and native multimodal vision via jointly trained DeepSeek-ViT embeddi

DeepSeek V4.1-Flash, released September 10, 2026 with MIT-licensed weights, is engineered around the observation that long-horizon agents are input-heavy workloads where prefill compute and KV-cache memory—not raw intelligence—are the primary cost bottleneck. The model splits its 40 Transformer layers into a 20-layer c Key verified specifications include a 552B MoE backbone across 40 layers, 890 bytes per token KV cache (one-quarter of V4-Flash), $0.003 per million cached input tokens at off-peak pricing, and a 1M token context via YaRN ×16. The architecture combines Compressed Sparse Attention 2 (CSA2) with three static attention mo

DeepSeek released DeepSeek-V4.1 Flash on September 10, 2026, describing it as the smallest model in a new architecture family that now natively supports visual understanding. The model uses a 552 billion parameter Mixture of Experts design with a new Causal Encoder-Decoder architecture, activating just 8 billion parame The article includes a comprehensive benchmark comparison table showing V4.1-Flash's performance against V4-Pro 0813, V4-Flash 0731, GLM 5.3, Kimi K3, GPT 5.6-Sol, and Claude Opus 5 across evaluations including GPQA Diamond (90.9), Terminal-Bench 2.1 (90.6), Terminal-Bench 4.0 (31.2), DeepSWE v1.1 (74.2), Codeforces ra