DeepSeek's V4.1-Flash Slashes Agent Memory Costs, Edges Out Opus 5 on Coding Benchmark
The MIT-licensed model packs 552 billion parameters but activates only 16 billion per token, cutting KV cache memory to a quarter of its predecessor's footprint.
DeepSeek released V4.1-Flash, a multimodal model that shrinks the KV cache memory required to run AI agents to a quarter of what its predecessor demanded. The Chinese lab published the weights under the MIT license, a permissive terms set that allows commercial use and modification without royalty payments. The move lands directly on the cost problem that has kept long-running agent workloads expensive to operate at scale.
The architecture is a mixture-of-experts design: 552 billion total parameters, but only 16 billion active for any given token. That sparsity is what keeps inference cheap despite the headline parameter count, since most of the network sits idle during each forward pass. DeepSeek says the KV cache reduction is the bigger practical win, because agent frameworks accumulate context across many tool calls and the cache grows with every turn.
On DeepSWE, a coding benchmark that tests software engineering tasks, V4.1-Flash narrowly outperformed both Opus 5 and GPT-5.6 Sol. The margin is thin enough that it may not hold across repeated runs or different harnesses, and benchmark results for agentic coding have proven sensitive to scaffolding and prompt format. Still, a model with 16 billion active parameters landing ahead of frontier closed models on coding is a result competitors will have to answer.
The KV cache cut matters most for the economics of agent deployment. Serving a long agent session means holding attention keys and values for the entire conversation history in memory, and that footprint scales with context length. Cutting it to a quarter means either four times the concurrent sessions on the same hardware or the same session count on cheaper GPUs, which changes the break-even point for anyone running coding agents or research assistants in production.
Multimodal capability plus an MIT license also lowers the barrier for teams that need to fine-tune or self-host rather than pay per-token API fees. DeepSeek has been building a pattern of releasing capable weights openly while Western labs keep their best models behind APIs, and V4.1-Flash extends that strategy into the agent tooling layer specifically. For startups building on top of agent frameworks, the option to run a frontier-adjacent model on their own infrastructure is a real alternative to renting one.
What to watch is whether the benchmark lead survives independent testing and whether the memory savings hold up in real agent loops rather than controlled evaluations. DeepSeek has not published throughput or latency figures against its own V4, and those numbers will decide whether V4.1-Flash is a genuine cost reduction or a memory optimization that shifts the bottleneck elsewhere. The weights are available now, so the answer should come from users rather than the lab's own charts.