← October 17, 2026 briefing · added in the 00:00 KST update
DeepSeek V4.1 Flash 1017 follow-up: company's first CED (Causal Encoder-Decoder) architecture, native image understanding; 552B total, 8B input/16B output active (community reports)
Announced October 17, 2026
Verified (fact)
According to a GeekNews post, DeepSeek V4.1 Flash 1017, released on Hugging Face, is a sparse MoE model and DeepSeek's first to use a CED (Causal Encoder-Decoder) architecture. Visual and text embeddings are trained jointly from the first stage of pretraining, giving it native image understanding. It has 552 billion total parameters, with 8 billion active when processing input and 16 billion active when generating output, an asymmetric design in which input and output active sizes differ. This architecture information is new, adding to the Terminal-Bench figures covered in an earlier briefing.
Why it matters (interpretation)
In open-weight MoE, where decoder-only architectures dominate, this is a case of using an encoder-decoder architecture with separately set active parameters for input and output. It may be a design to cut inference costs for agent tasks with many input tokens, so it is worth watching whether other labs follow.
Confidence medium · official source pending
Sources
- Community DeepSeek V4.1 Flash 1017 - CED 아키텍처와 네이티브 이미지 이해를 도입한 MoE 모델 GeekNews
Story thread
- 2026-10-17 DeepSeek releases 'V4.1-Flash-1017' on Hugging Face; Terminal-Bench 4.0 rises from 31.2 to 42.3 (community reports)
- 2026-10-17 DeepSeek V4.1 Flash 1017 follow-up: company's first CED (Causal Encoder-Decoder) architecture, native image understanding; 552B total, 8B input/16B output active (community reports)
- 2026-10-18 DeepSeek V4.1-Flash-1017 follow-up: official release confirmed with MIT-licensed weights, rolled out to web, app and API at unchanged prices (report)
- 2026-10-19 DeepSeek V4.1-Flash-1017 follow-up: independent Terminal-Bench 4.0 scores of 34% and 31.8% trail the model card's 42.3% (community reports)
More about DeepSeek
-
Follow-up to DeepSeek API tool rollout: no separate fee for model-initiated web search and fetch; direct calls still billed per request (Pandaily)
According to Pandaily, DeepSeek has dropped the separate fees it charged for web searches and web page fetches that the model runs on...
-
Alibaba's workplace AI 'QwenWork' adds third-party models including Kimi, GLM and DeepSeek, with access control by department, group or user (Pandaily)
According to Pandaily, Alibaba has expanded the models available in QwenWork, its enterprise work tool. Besides its own Qwen models,...
-
Users report daily usage caps and input cut to ~40,000 characters on DeepSeek's free app and web chat; DeepSeek has not confirmed (Pandaily)
Pandaily relayed user reports that DeepSeek's free app and web chat now have daily usage limits, and that the maximum input per message...
-
LangChain ships new Mistral, Groq, xAI and DeepSeek integration packages that route chat model calls through the LangSmith gateway
LangChain released langchain-mistralai 1.2.0, langchain-groq 1.2.0, langchain-xai 1.4.0 and langchain-deepseek 1.2.0. All four packages...
-
Usage-limit code pointing to a 'Coding Plan' found in DeepSeek's web app; no subscription, pricing or launch date announced (Pandaily)
According to Pandaily, developers found code in DeepSeek's web chat that manages usage limits and refers to a 'Coding Plan'. DeepSeek...