DeepSeek has releasedDeepSeek V4.1 Flash, its newest AI model designed to deliver stronger performance while dramatically improving speed and cost efficiency.
The model launched on September 10 and is the smallest model in DeepSeek's new architecture family. It comes withnative image understanding, higher throughput and a new architecture designed to make AI agents more efficient.
A smaller model with a big ambition
DeepSeek V4.1 Flash uses a552-billion-parameter Mixture-of-Experts architecture, but only a fraction of those parameters are activated for each token.
According to DeepSeek, the model activates around8 billion parameters during input processing and 16 billion during output generation. This architecture is designed to reduce the computing resources required to process large amounts of information.
The model also supports contexts of up to1 million tokensand can process both text and images. Its model weights have been released under the MIT license, giving developers more flexibility to experiment with and deploy the technology.
DeepSeek says it beats its own V4 Pro
One of the most interesting parts of the announcement is that DeepSeek says V4.1 Flash hassurpassed V4 Pro in performance, cost, speed and total runtimebased on extensive testing.
As a result, DeepSeek is now phasing out V4 Pro.
StartingSeptember 14 at 04:00 UTC, requests made to thedeepseek-v4-proAPI will automatically be routed to V4.1 Flash and charged at the newer Flash pricing until V4.1 Pro is released.
That's an unusual move in the AI industry: a newer, cheaper model is effectively replacing a more expensive flagship model.
Why developers should pay attention
The biggest attraction may not be raw benchmark scores.
It'scost efficiency.
DeepSeek has reduced its API pricing alongside the V4.1 Flash release, while the model is specifically designed for workloads involving AI agents, coding and long-running tasks.
That could make it attractive to developers building AI-powered applications where models have to process large amounts of input repeatedly.
DeepSeek's published benchmarks also show strong results across reasoning, coding and agent-related tasks, although these figures are the company's own reported results and should be interpreted alongside independent testing.
Another challenger in the AI race
The timing is significant.
OpenAI has recently launchedGPT-6 Astra, Google has releasedGemini 3.8 Flash, and Anthropic continues to compete aggressively in advanced AI.
DeepSeek is taking a different approach:make powerful AI cheaper and more accessible while releasing model weights openly.
If that strategy continues, the competition may increasingly be about more than simply building the smartest model.
It could become a battle overwho can deliver the most useful AI at the lowest cost.
And DeepSeek V4.1 Flash has just entered that race.
