● Covered by 3 sources · 4 reportsMedium impact4 neutral
DeepSeek Releases V4.1 Flash 1017 Model with New CED Architecture and Asymmetric Parameter Activation
Key points
- DeepSeek V4.1 Flash 1017 is a sparse mixture-of-experts model.
- It uses a Causal Encoder-Decoder (CED) architecture.
- The model activates 8B parameters on input and 16B on output from 552B total.
- Native image understanding is integrated from pre-training.
- DeepSeek V4.1 Flash 1017 (Max) scores 45 on the Artificial Analysis Intelligence Index.
- The model processes 227 tokens per second.
- Input tokens cost $0.30 per 1M.
- Output tokens cost $1.20 per 1M.
- The model has a 1M tokens context window.
- It supports text and image input, and outputs text.
- DeepSeek V4.1 Flash 1017 was released on October 17.
- DeepSeek V4.1 Flash 1017 cuts task runtime and output length.
- DeepSeek V4.1 Flash 1017 scores 42.3 on Terminal-Bench 4.0.
- DeepSeek V4.1 Flash 1017 is available via DeepSeek's API, OpenRouter, or Hugging Face.
- Claude Haiku 5.5 was released on October 7.
- Claude Haiku 5.5 costs $0.10 per million input tokens.
- Claude Haiku 5.5 costs $0.50 per million output tokens.
- Claude Haiku 5.5 scores 39.2 on Terminal-Bench 4.0.
- DeepSeek V4.1 Flash 1017 is an open-weight model.
- DeepSeek V4.1 Flash 1017 is suitable for open-ended tasks.
- DeepSeek V4.1 Flash 1017 can be hallucination-prone.
- DeepSeek V4.1 Flash 1017 activates 16B parameters during runs.
- DeepSeek V4.1 Flash 1017 slots between GPT-5.6 Sol and Claude Opus 5.
DeepSeek V4.1 Flash 1017 Release
DeepSeek has released V4.1 Flash 1017, which is the October 17, 2026 snapshot of its V4.1 Flash model. This model is served under the deepseek-flash API model name and represents an advancement in the company's AI offerings.
New Architecture and Parameter Activation
V4.1 Flash is a sparse mixture-of-experts model, marking the first to be built on DeepSeek's Causal Encoder-Decoder (CED) architecture. It features an asymmetric parameter activation, using 8 billion parameters on input and 16 billion on output, from a total backbone of 552 billion parameters. This design aims to maintain low per-token compute relative to the model's overall size.
Integrated Image Understanding
A key feature of the V4.1 Flash architecture is its native image understanding capability. Visual and text embeddings are trained jointly from the beginning of pre-training, a departure from previous experimental models like V4 Flash Vision Exp where such capabilities were added post-training.
Target Applications and Cost Efficiency
The model is designed for applications such as coding, terminal, and computer-use agents, as well as long-horizon tasks requiring multiple steps and extensive context analysis. DeepSeek positions V4.1 Flash as the cost-efficient tier within its V4.1 family. It reportedly surpasses the performance, speed, and task completion times of the earlier V4 Pro model.
Memory Optimization
V4.1 Flash incorporates compressed KV caching, which reduces cache memory usage to approximately one-quarter of the previous Flash generation. This reduction is intended to lower costs for agentic workloads.
Updates
- DeepSeek V4.1 Flash 1017 is an open-weight model.
- DeepSeek V4.1 Flash 1017 is suitable for open-ended tasks.
- DeepSeek V4.1 Flash 1017 can be hallucination-prone.
- DeepSeek V4.1 Flash 1017 activates 16B parameters during runs.
- DeepSeek V4.1 Flash 1017 slots between GPT-5.6 Sol and Claude Opus 5.
- DeepSeek V4.1 Flash 1017 was released on October 17.
- DeepSeek V4.1 Flash 1017 cuts task runtime and output length.
- DeepSeek V4.1 Flash 1017 scores 42.3 on Terminal-Bench 4.0.
- DeepSeek V4.1 Flash 1017 is available via DeepSeek's API, OpenRouter, or Hugging Face.
- Claude Haiku 5.5 was released on October 7.
- Claude Haiku 5.5 costs $0.10 per million input tokens.
- Claude Haiku 5.5 costs $0.50 per million output tokens.
- Claude Haiku 5.5 scores 39.2 on Terminal-Bench 4.0.
- DeepSeek V4.1 Flash 1017 (Max) scores 45 on the Artificial Analysis Intelligence Index.
- The model processes 227 tokens per second.
- Input tokens cost $0.30 per 1M.
- Output tokens cost $1.20 per 1M.
- The model has a 1M tokens context window.
- It supports text and image input, and outputs text.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub Kilo-Org/kilocode GitHub deepseek-ai/DeepSeek-V2 GitHub meta-llama/llama-models GitHub QwenLM/Qwen GitHub deepseek-ai/deepseek-LLM GitHub deepseek-ai/DeepSeek-Coder-V2How outlets covered it
Deepseek updated its V4.1 Flash open-weight AI model to revision 1017, significantly cutting hardware requirements and operational costs. This update positions the model as a more efficient alternative for open-ended tasks, despite potential hallucination issues.
DeepSeek V4.1-Flash-1017, an update released October 17, was compared against Anthropic's Claude Haiku 5.5, released October 7, for accuracy and cost. The comparison found DeepSeek's model to be more accurate in specific tests and potentially cheaper for longer prompts, despite Haiku's lower per-token rates for short prompts.
An analysis of the DeepSeek V4.1 Flash 1017 (Max) model shows it ranks highly in intelligence among comparable open-weight models, scoring 45 on the Artificial Analysis Intelligence Index. The model is also noted for its speed at 227 tokens per second and moderate pricing, though it is very verbose in its output.
DeepSeek released V4.1 Flash 1017, a new sparse mixture-of-experts model built on its Causal Encoder-Decoder (CED) architecture. This model activates 8B parameters on input and 16B on output from a 552B-parameter backbone, and includes native image understanding, aiming for cost-efficient performance in agentic workloads.