AI research laboratory DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, making its weights available under an open MIT license on Hugging Face and ModelScope. The underlying architecture relies on a Mixture-of-Experts (MoE) system encompassing 552 billion total parameters. The model employs an asymmetric causal encoder-decoder structure where 8 billion parameters are active during input processing and 16 billion parameters are activated during output generation.
A central architectural objective of the update is handling massive context windows efficiently. DeepSeek-V4.1-Flash supports an input length of up to one million tokens. By combining an asymmetric structure with compressed sparse attention, the memory footprint of the key-value cache (KV cache) is cut to roughly one-fourth compared to earlier designs. This reduction makes running the model on standard GPU configurations, such as four Nvidia H200 accelerators, significantly more viable.
The model demonstrated gains in specialized software engineering evaluations. In the DeepSWE v1.1 benchmark, the score climbed from 74.2 points to 77.0 points. On Terminal-Bench 4.0, which evaluates command-line tasks, the system recorded 42.3 points. DeepSeek positions the release primarily as an engine for developer agents, multistage coding workflows, game creation pipelines via Godot and automated 3D scene generation.
Alongside the model weights, DeepSeek launched a dedicated Search and Fetch API to supply agentic systems with web data. The service carries a standard price of $0.75 per 1,000 queries. DeepSeek announced that access remains free until October 31, 2026, for integrations in coding tools such as Claude Code and Codex.
This release arrives during shifting alliances in the software-generation ecosystem. Replit CEO Amjad Masad recently revealed at TechCrunch Disrupt in mid-October that Replit Agent switched its default backend from Anthropic to OpenAI due to benchmark performance and cost efficiency. As proprietary providers compete on price, open-weight architectures like DeepSeek give engineers flexible alternatives to host agent frameworks locally and lower operating expenses.