DeepSeek launches low-cost V4.1-Flash model that outperforms leading AI systems in key tests
Chinese AI firm DeepSeek introduced its budget V4.1-Flash model, which beats several top-tier models on benchmarks while offering lower token pricing. The model also supports direct image input and can be run on user hardware under an MIT licence.
DeepSeek unveiled V4.1-Flash, a budget AI model that combines a new causal encoder-decoder architecture with a 552-billion-parameter mixture-of-experts design, activating only a fraction of its parameters per request. In internal benchmarks, the model achieved 90.6 points on Terminal-Bench 2.1 and 54.8 on AutomationBench, outpacing Anthropic’s Opus-5.0 and OpenAI’s GPT-5.6 Sol. Pricing is set at $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours, markedly cheaper than rivals.
V4.1-Flash can process both images and text, enabling direct analysis of receipts, charts or screenshots. The model is available for download under an MIT licence, allowing users to run it on their own hardware. Concurrently, DeepSeek is reportedly preparing a large AI data centre in Ulanqab, Inner Mongolia, featuring up to 160,000 Huawei Ascend 950DT inference processors, a move intended to reduce dependence on US-made chips.
