DeepSeek-V4.1-Flash-1017

DeepSeek-V4.1-Flash-1017 is an open multimodal model from DeepSeek-AI built on the basis of Causal Encoder-Decoder (CED) and Compressed Sparse Attention 2 (CSA2) under the MIT license. The model uses 763B parameters, of which 552B are in Mixture-of-Experts, 196B in the Engram associative memory module, and about 14B in the DSpark generation acceleration module. The model accepts text and images as input, works with a context of up to 1M tokens, and activates 16B parameters per token during response generation and only 8B when processing the prompt. The -1017 suffix is a point update to the September DeepSeek-V4.1-Flash: the architecture and parameters are unchanged, but quality has noticeably increased.

On benchmarks, version 1017 noticeably outperforms DeepSeek-V4.1-Flash in all areas and in a number of tests takes first place among top frontier models. On DeepSWE v1.1, the model scores 77.0 versus 74.2 for DeepSeek-V4.1-Flash, 74.0 for Claude Opus 5, 73.0 for GPT-5.6 Sol, 67.5 for Kimi K3, and 66.9 for GLM-5.3. On Automation-Bench v1.0.6, the result is 60.5 versus 54.8 for the previous version, 50.3 for Claude Opus 5, 48.8 for GLM-5.3, 46.7 for Kimi K3, and 45.8 for GPT-5.6 Sol. The model also leads on Agents’ Last Exam (35.4) and Apex Bench v1.1 (72.1), outperforming the listed top models. On Terminal-Bench 4.0, the score rose from 31.2 to 42.3; on HLE without tools, from 36.8 to 42.6. Strong results are additionally noted on Chartography (83.0) and BabyVision (92.5).

Thus, 1017 retains all the architectural advantages of CED and CSA2, including KV cache savings and nearly constant decode-FLOPs, but thanks to improved post-training and optimizations it delivers higher results, making reasoning more meaningful and faster, and answers shorter and more precise. It is this combination—better metrics with less generation time and volume—that makes 1017 more advantageous for industrial deployment than DeepSeek-V4.1-Flash. At the same time, the model is aimed at a wide range of tasks, including various agentic scenarios, software development, office automation, and, of course, conversational assistants in various domains.

Announce Date: 17.10.2026
Parameters: 764B
Experts: 384
Activated at inference: 16B
Context: 1049K
Layers: 40, using full attention: 4
Attention Type: Causal Encoder‑Decoder (CED) и Compressed Sparse Attention 2 (CSA2)
Developer: DeepSeek
Transformers Version: 5.6.0
vLLM Version: >=0.31.0
License: MIT

Use our pre-built public endpoints for free to test inference and explore DeepSeek-V4.1-Flash-1017 capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.