Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
DeepSeek released V4.1-Flash, a 552B-parameter model that uses only 8B active parameters for input processing and 16B for output, cutting KV cache requirements to 1/4 the memory and 1/8 the storage of the previous generation. The company is retiring V4-Pro and routing all its traffic to V4.1-Flash at lower prices starting September 14, 2026.
- V4.1-Flash outperforms V4-Pro on benchmarks while using asymmetric encoder-decoder architecture that dramatically reduces active parameters and cache overhead
- KV cache compression cuts memory by 75% and storage by 87.5%, directly lowering inference costs for agents and long-running tasks
- Pricing drops on September 10, 2026, with off-peak rates at 50% of peak rates; V4-Pro requests automatically migrate to V4.1-Flash at the new lower rates
DeepSeek released V4.1-Flash, a new model built on an asymmetric architecture that splits computation between input and output processing. The model has 552 billion total parameters but only activates 8 billion for inputs and 16 billion for outputs, cutting inference costs dramatically. This design comes from new pretraining methods and scaled-up reinforcement learning that apparently pushed performance past their flagship V4-Pro model across multiple benchmarks. The KV cache—the memory overhead that compounds costs in long conversations—shrinks to one-quarter the size of the previous generation for RAM and one-eighth for storage, which matters because cache hits often dominate API bills for agents running repeated queries.
The practical shift here is that V4-Pro is being phased out. Starting September 14, 2026, all requests to the V4-Pro endpoint automatically route to V4.1-Flash at V4.1-Flash pricing. Independent tests apparently showed V4.1-Flash winning on performance, cost, speed, and total runtime combined. The model now ships with native image understanding built in, and DeepSeek's API pricing drops on September 10 to reflect the efficiency gains. They're keeping the existing peak/off-peak structure where off-peak costs half as much, which rewards teams that can schedule batch work during cheaper hours.
DeepSeek is also signaling openness to custom deployments. They're explicitly inviting conversations about large-scale setups using 2,000+ GPUs with storage clusters, suggesting they want to support companies building private inference infrastructure. Official partners like WorkBuddy and CodeBuddy already support V4.1-Flash, so integration paths exist for teams already embedded in those ecosystems. The company plans to work with the open-source community on inference optimization, though specifics on that collaboration are light.