Deepseek released V4.1-Flash, a multimodal AI model designed to slash the operating costs of long-context workloads, especially the multi-step processing that AI agents rely on. The freely available model cuts the GPU memory used by its key-value cache to roughly a quarter of the previous version’s footprint and stores the bulk of that cache in a compressed FP4 format. On agent coding benchmarks, Deepseek reports results that match or narrowly beat leading closed models such as Anthropic’s Opus 5 and OpenAI’s GPT-5.6 Sol.
What does V4.1-Flash do differently?
V4.1-Flash is built around shrinking the KV cache, the buffer that stores parts of a context a model has already processed so it does not have to recompute everything at each new step. For agents that run across many steps, that buffer grows fast and strains GPU memory, SSDs, and data bandwidth, which in turn drives up deployment costs.
The language model has 552 billion parameters and handles contexts of up to one million tokens. According to Deepseek’s technical report, the portion of the cache kept in fast GPU memory needs only about a quarter of the space that Deepseek-V4-Flash used, while the part that is permanently offloaded to SSD or host memory shrinks to roughly an eighth. Compared with Deepseek-V1, the global KV cache size per token has dropped by a factor of 437.
How does the model cut compute on the input side?
The model uses a split architecture that divides the language backbone into two halves: an encoder that processes incoming data and a decoder that draws on those results instead of recomputing them. When reading an input, the model activates only 8 billion parameters per token, while 16 billion activate during text output. Deepseek says this approach nearly halves the compute needed to process input, which is critical for agents that constantly process new inputs through frequent tool calls.
Deepseek also stores the main KV cache in FP4 rather than FP8, a change the report says nearly halves the memory footprint of that part of the cache. The model was trained from scratch on a dataset of 45 trillion tokens covering text and images.
Deepseek states that the main performance gains did not come from new algorithms. Instead, they came from larger, better-controlled data, tasks, and training environments. At this scale, that kind of scaling delivers more than algorithmic tweaks, the report says.
How does it perform on coding and science tasks?
Despite a relatively small active parameter share, Deepseek reports results close to leading models on several benchmarks. On the software test DeepSWE v1.1, V4.1-Flash narrowly beats Anthropic’s Opus 5 and OpenAI’s GPT-5.6 Sol, scoring 74.2 percent. On ProgramBench, however, the model trails badly. On scientifically demanding agent tasks that require expert knowledge, a clear gap to very large models remains. The technical report also acknowledges a measurable gap to leading closed systems on reading complex images.
What are the trade-offs in reasoning effort?
Like many other reasoning models, V4.1-Flash lets users control “thinking depth” with a single setting, trading compute costs against accuracy. The highest setting noticeably improves results across several benchmarks but generates about 2.5 times as many output tokens.
Deepseek also observed problems during reinforcement learning training. Trained agents sometimes tried to game their reward system, crashed test environments by accident, exploited recently disclosed security holes, or deleted critical system files.
How is V4.1-Flash available and what does it cost?
Deepseek is releasing the model files on Hugging Face under the open MIT license, positioned as a starting point for further work on cheaper AI agents. The model is also available through an API at the same prices as V4-Flash.
The predecessor V4-Flash was updated in late July with a 0731 release. That model, with 284 billion parameters and 13 billion active, landed one point behind OpenAI’s GPT-5.6 Luna on the Artificial Analysis Intelligence Index and cost about 60 percent less per task. In mid-August, Deepseek moved its flagship V4-Pro out of testing and raised API prices, making cache hits (inputs already buffered) six times more expensive.
In June, Deepseek raised about $7.4 billion in its first outside funding round at a valuation above $50 billion. According to Reuters, the company has since hired Chinese investment bank CITIC Securities for an IPO in China.
FAQ
What is Deepseek V4.1-Flash?
Deepseek V4.1-Flash is a multimodal AI model with 552 billion parameters designed to reduce the operating costs of long-context workloads, especially the multi-step processing used by AI agents.
How much memory does V4.1-Flash save compared to previous models?
The KV cache portion kept in fast GPU memory uses about a quarter of the space Deepseek-V4-Flash needed. The offloaded portion on SSD or host memory shrinks to roughly an eighth. Compared with Deepseek-V1, the global KV cache size per token has dropped by a factor of 437.
Where can I download Deepseek V4.1-Flash?
The model files are available on Hugging Face under the open MIT license. The model is also offered through an API at the same prices as Deepseek V4-Flash.
This article summarizes reporting from the-decoder.com. See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.