DeepSeek V4.1 Flash is now available on Verda Inference, at $0.15 per million input tokens, $0.006 for cached input and $0.60 for output. That’s about 50% less than most providers charge.* It’s DeepSeek's capable, efficient model for everyday coding and agents, with a 1M-token context, image input, and thinking you can switch on or off.
It joins Moonshot AI’s Kimi K3.1 Pro on Verda Inference, so you can choose the right model for each task through a single API. Below, we cover what makes DeepSeek V4.1 Flash efficient, and how we serve it.
Built for the tasks you run every day
DeepSeek V4.1 Flash is built to do a lot of work for little compute. It’s a 552B-parameter Mixture-of-Experts model that activates only 8B parameters while reading input and 16B while writing output. Agents spend most of their time reading, so that keeps them affordable.
It also keeps long contexts light. To work with a long context, a model keeps a memory of everything it has read, called the KV cache. DeepSeek V4.1 Flash keeps that memory small, at 890 bytes per token, so long sessions take up less GPU memory and stay fast.
And you can choose how hard it thinks. Switch thinking off for fast answers, or on for harder problems, with up to 384K tokens of output.
Served at the efficiency of DeepSeek’s own engine
The same model can perform very differently depending on how it’s run. The software and setup behind it decide how fast answers arrive, how many requests each GPU can handle, and how well performance holds up as traffic grows.
Verda Inference is built on SGLang, the open-source serving engine. Before launch, we tuned our deployment of DeepSeek V4.1 Flash and benchmarked it against the SGLang team’s published results, which put SGLang on par with DeepInfer, DeepSeek’s own engine.
Models do two kinds of work: reading the prompt, and writing the answer. On both, our results matched theirs. Prompts were read within about 1% of the published results, and answers were written 1 to 6% faster at every load level, with output peaking at about 29,300 tokens per second per GPU.**
For you, that means DeepSeek V4.1 Flash keeps up as your workload grows, whether you’re reviewing every pull request in a codebase, or running every step of a long agent task.
In an upcoming post, we’ll dive deep into the engineering behind these results.
Get started
Verda Inference is OpenAI compatible, so you can start with the code and tools you already use.
- Enable DeepSeek V4.1 Flash in the Verda console.
- Create your API key.
- Point your OpenAI client at
https://inference.verda.com/v1, with the model IDdeepseek-4-1-flash-1017. - Choose thinking or non-thinking mode with
reasoning_effort. Thinking is off by default.
See the model page and the docs for details, or set it up in coding agents like OpenCode, Pi Coder and Codex with our ready-made configs.
*Compared with most providers’ list prices, as of November 2026. Prices subject to change.
**Based on an internal engine benchmark using the SGLang team’s published method, under simulated conditions. Results may not reflect performance in production.
