According to DeepSeek, the updated model has ‘substantial improvements’ across reasoning, coding, agentic tasks and visual benchmarks.
Chinese AI company DeepSeek has released an update of its V4.1-Flash model, which it said has stronger performance, faster task completion and shorter outputs than the previous version of the AI model.
The company announced the new update – V4.1-Flash-1017 – in a blogpost on Saturday (17 October), which is live on DeepSeek’s website, app and API.
According to a technical report that DeepSeek published on AI platform Hugging Face, the updated model has “substantial improvements” across reasoning, coding, agentic tasks and visual benchmarks, and allows users to complete tasks faster and at a lower cost.
For example, DeepSeek recorded a score of 77 on software engineering benchmark DeepSWE for the updated model. In comparison, the previous version of the model scored 74.2, while OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Opus 5 scored 73 and 75 respectively.
DeepSeek also claimed that V4.1-Flash-1017 particularly enhances capabilities in video game creation and 3D modelling, thanks to the use of reinforcement learning with self-feedback to “improve end-to-end project development”.
The start-up showcased some examples of these enhanced capabilities being utilised in its blogpost, such as through a 3D model of China’s Forbidden City, a web-based racing game and a recreation of the game Portal 2 built with open-source engine Godot.
Saturday’s update builds on DeepSeek’s V4.1-Flash AI model, which the company released in September and marketed as a “smarter, faster, more efficient” model. The release of V4.1-Flash, which operated on lowered token prices, caused share prices to drop for a number of DeepSeek’s rivals.
The global race for AI dominance has been accelerating in recent months, with open-weight models becoming a crucial influence.
Open-weight models are considered necessary for AI leadership, as being able to tweak a model means organisations, from start-ups to public institutions, can build on advanced models without training one from scratch.
While Chinese companies predominantly lead the open-weights space, various global rivals have been working on catching up.
Last week, US tech giant Meta unveiled Muse Spark Open, which the company called the “most capable” US open-weight model to date.
The multimodal model has 816bn total parameters, with 42bn active parameters and a 1m-token context window. According to the company, it is built for coding, agentic work and other complex, multi-step reasoning tasks.
Meanwhile, earlier this month, French AI company Mistral announced its latest open-weight AI model Mistral Large 4, which it said “significantly outperforms” any open-weight model developed in the US or Europe.
Known as ‘Le Chonk’, ML4 is a 1trn-parameter natively multimodal model with 49bn active parameters, and Mistral’s “largest and most capable model to date”.
Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.