DeepSeek announced Thursday the official release of V4.1 Flash — the smallest model in its new architecture family, arriving with native multimodal visual understanding. The positioning is blunt: greater capability ceiling, faster inference, higher throughput, and a design meant to scale up to much larger models.
V4.1 Flash is a 552-billion-parameter mixture-of-experts (MoE) model built on a new Causal Encoder–Decoder architecture. Its most distinctive feature is asymmetry: just 8B active parameters for input and 16B for output. That keeps cost significantly below known models of comparable size — more intelligence, less expense. Combined with new pretraining methods and larger-scale reinforcement-learning post-training, V4.1 Flash delivers benchmark results ahead of flagship models, including DeepSeek's own V4 Pro.
KV-cache compression is the other headline. Compared with the previous generation, V4.1 Flash's KV cache needs just one quarter of the HBM and one eighth of the SSD storage. Cache-hit charges often account for a large share of agent costs, so shrinking the cache directly slashes the most painful line item in agentic workloads. The model is live on the DeepSeek API today — set your model to deepseek-flash to call it. V4 Flash and V4 Flash Vision Exp are retired, with their old names temporarily routing to the new model for compatibility.
On pricing, DeepSeek continues its peak/off-peak scheme: off-peak rates are 50% of peak, encouraging developers to schedule flexible workloads at low-traffic hours. Multiple third-party tests put V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime, so the company is phasing V4 Pro out — starting at 04:00 UTC on September 14, all deepseek-v4-pro requests will route to V4.1 Flash at V4.1 Flash rates, until V4.1 Pro launches. Official partners WorkBuddy (including CodeBuddy) and OpenCode fully support V4.1 Flash now.
Open source remains DeepSeek's foundation. V4.1 Flash weights and the technical report are on Hugging Face under an MIT license, with a 1M-token context window. The company says it will work closely with the open-source community on inference support and explore more deployment options — and adds, pointedly, that teams planning large-scale deployments with 2,000 GPUs plus a storage cluster should get in touch.
The combination of small size, big intelligence, and lower cost keeps raising the value ceiling of frontier AI.