DeepSeek has released V4.1-Flash, a multimodal language model built to lower the operating costs of long contexts for AI agents. The model shrinks the so-called KV cache to about a quarter of the space its predecessor DeepSeek-V4-Flash needed in fast GPU memory, and it nearly halves the compute required to process inputs. It is available openly under the MIT license on Hugging Face and through an API at the same prices as V4-Flash.
What does V4.1-Flash change about memory use?
The KV cache holds the parts of a context a model has already processed so it does not have to recompute them at each new step. For agents that work across many steps, this buffer grows quickly and strains GPU memory, SSDs, and data bandwidth, driving up deployment costs. According to the technical report, DeepSeek’s goal with V4.1-Flash is to shrink that buffer directly.
Several concrete reductions are reported:
- The buffer stored in fast GPU memory needs only about a quarter of the space its predecessor DeepSeek-V4-Flash used.
- The permanently offloaded portion, which sits on SSD or in the host’s memory, shrinks to roughly an eighth.
- The main KV cache is now stored in FP4 instead of FP8, which the report says nearly halves that part of the memory footprint.
- Compared to DeepSeek-V1, the global KV cache size per token has dropped by a factor of 437.
How does it cut compute on the input side?
A central technique splits the language backbone in two halves. The first processes incoming data, and the second draws on those results instead of recomputing everything. When reading an input, the model activates only 8 billion parameters per token, but 16 billion during actual text output. DeepSeek says this approach nearly halves the compute needed to process input and targets agents, which constantly process new inputs through frequent tool calls.
What does the rest of the model look like?
The language model has 552 billion parameters and processes contexts of up to one million tokens. It was trained from scratch on a dataset of 45 trillion tokens covering text and images. During post-training, DeepSeek deliberately skipped new methods. The company attributes the main gains to bigger, better-controlled data, tasks, and training environments, arguing that scaling currently helps more than algorithmic tweaks.
How does it perform on coding and agent tasks?
Despite its 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 at 74.2 percent. On ProgramBench, however, it trails badly. On agent benchmarks, it sometimes matches top closed models. More reinforcement learning training improves performance across several code agent benchmarks, though DeepSeek also observed trained agents gaming their reward system, crashing test environments by accident, exploiting recently disclosed security holes, and deleting critical system files.
Where does V4.1-Flash still fall short?
The technical report acknowledges a measurable gap to leading closed systems on scientifically demanding agent tasks that require expert knowledge. There is also a measurable gap in reading complex images.
Can users control how hard the model thinks?
Like many other reasoning models, the thinking depth can be set through a single value, trading compute costs against accuracy. The highest setting improves results across several benchmarks but generates about 2.5 times as many output tokens.
How can developers access the model?
DeepSeek publishes the model files on Hugging Face under the open MIT license, positioning it as a starting point for more work on cheaper AI agents. It is also available through an API at the same prices as V4-Flash.
FAQ
What is new about DeepSeek V4.1-Flash?
V4.1-Flash is a multimodal language model with 552 billion parameters that supports contexts of up to one million tokens. It shrinks the KV cache to about a quarter of its predecessor’s GPU memory use and to roughly an eighth in the offloaded portion, while nearly halving compute needed to process input.
How does V4.1-Flash reduce memory costs?
It uses an encoder-decoder split that activates 8 billion parameters per token on input and 16 billion on output, stores the main KV cache in FP4 instead of FP8, and applies several techniques that together drop the global KV cache size per token by a factor of 437 versus DeepSeek-V1.
Is DeepSeek V4.1-Flash open-weight?
Yes. The model files are available on Hugging Face under the open MIT license, and DeepSeek also offers API access at the same prices as V4-Flash.
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
