DeepSeek V4.1 Flash 1017 Ultrafast

• 🧠 Multimodal Mixture-of-Experts, 552B backbone, 8B/16B active parameters
• 📏 One-million-token context window for long documents and agents
• 🆕 Causal Encoder-Decoder design, a first for DeepSeek
• 👁️ Natively reads images and text, outputs text
• ⚡ Speed-oriented FP8 serving of the October 17 snapshot
• 🔧 Function calling, web search, coding and agentic workflows
• 🔒 Open weights released under the MIT license

DeepSeek

DeepSeek is a Chinese artificial intelligence company specializing in large language model development, founded in July 2023 by Liang Wenfeng. Based in Hangzhou, Zhejiang, the company is backed by High-Flyer, a prominent Chinese hedge fund also co-founded by…

DeepSeek V4.1 Flash 1017 Ultrafast is a throughput-oriented, FP8-quantized deployment of the October 17, 2026 snapshot of DeepSeek's V4.1 Flash line. The underlying model is a multimodal Mixture-of-Experts system with a 552B-parameter backbone and support for contexts up to one million tokens, processing images and text natively and generating text autoregressively.

Architecturally it departs from earlier DeepSeek releases such as DeepSeek V4 Flash 0731 and DeepSeek V3.2: V4.1 Flash is the first DeepSeek model built on a Causal Encoder-Decoder layout — 40 layers split into a 20-layer causal encoder and a 20-layer decoder, with the decoder's global KV cache projected from final encoder states. That asymmetry activates only about 8B parameters per prompt token and 16B per output token, which DeepSeek describes as improving cost efficiency on input-heavy agentic workloads. A from-scratch vision encoder and MLP projector feed visual embeddings into the language stack from the start of pre-training.

Relative to the September DeepSeek V4.1 Flash launch checkpoint, the 1017 revision keeps the same architecture and model family naming. The same snapshot is catalogued as DeepSeek v4.1 Flash (1017); community variants include DeepSeek V4.1 Flash 1017 Uncensored. As reported on the NVIDIA NIM model card for the instruct model, it scores 90.9 Pass@1 on GPQA Diamond and 74.2 resolved on DeepSWE v1.1.