Skip to content

DeepSeek V4.1 Flash

DeepSeek's smaller V4.1 model: 552B parameters, 8B active on input and 16B on output, a 1M context, vision in and text out. Open weights.

Open DeepSeek V4.1 Flash
Bar chart comparing DeepSeek-V4.1-Flash with Kimi-K3, GLM-5.3, Opus 5 and GPT-5.6 Sol on Terminal-Bench, DeepSWE, CyberGym and Automation-Bench

deepseek/deepseek-v4.1-flash is DeepSeek V4.1 Flash, the smaller model in the V4.1 family. DeepSeek describes a 552B-parameter mixture of experts with a causal encoder-decoder: 8B parameters active while it reads, 16B while it writes. It takes text and images, answers in text, and the context window is 1,048,576 tokens. Reasoning effort is controllable. The weights are on Hugging Face under MIT. The release note and the tech report are DeepSeek's.

DeepSeek's own agentic chart puts V4.1 Flash at 30.0 on Terminal-Bench 3.0, 74.2 on DeepSWE v1.1, 88.1 on CyberGym, and 54.8 on Automation-Bench. Those bars are the vendor's comparison with Kimi-K3, GLM-5.3, Opus 5, and GPT-5.6 Sol.

Bar chart of DeepSeek V4.1 Flash against Kimi-K3, GLM-5.3, Opus 5 and GPT-5.6 Sol on Terminal-Bench 3.0, DeepSWE v1.1, CyberGym and Automation-Bench. DeepSeek published the chart.
Agentic scores from DeepSeek's V4.1 Flash model card.

Epoch AI scores in our catalog, separate from that chart: GPQA Diamond 89.8, OTIS Mock AIME 98.3, ARC-AGI-2 72.9, SciCode 51.9, SimpleQA Verified 42.9.

BenchmarkScoreSource
Terminal-Bench 3.030.0DeepSeek
DeepSWE v1.174.2DeepSeek
CyberGym88.1DeepSeek
Automation-Bench54.8DeepSeek
GPQA Diamond89.8Epoch AI
OTIS Mock AIME98.3Epoch AI
ARC-AGI-272.9Epoch AI
SciCode51.9Epoch AI
SimpleQA Verified42.9Epoch AI

The same card shows the KV cache shrinking across DeepSeek generations. DeepSeek says V4.1 Flash needs about a quarter of the previous generation's HBM for that cache, and about an eighth of the SSD. Cache-hit cost is a large share of a long agent run, so a smaller cache is the point of the architecture.

Bar chart of global KV cache bytes per token for DeepSeek V1, V3.2, V4-Flash and V4.1-Flash. Each generation is smaller.
KV cache bytes per token, from DeepSeek's model card.

Call it with your own key. AnyRouter adds no per-token charge on that path. Try it with the Vercel AI SDK; createAnyRouter() reads ANYROUTER_API_KEY:

generateText with the catalog id
import { createAnyRouter } from "@anyr/ai-sdk-provider"
import { generateText } from "ai"

const anyrouter = createAnyRouter()

const { text } = await generateText({
  model: anyrouter("deepseek/deepseek-v4.1-flash"),
  prompt: "Rank these three refactors by risk and say why.",
})
console.log(text)

Or launch a coding agent through the anyr CLI:

curl -fsSL https://anyr.dev/cli.sh | bash
anyr login
anyr claude --model "deepseek/deepseek-v4.1-flash"
curl -fsSL https://anyr.dev/cli.sh | bash
anyr login
anyr opencode --model "deepseek/deepseek-v4.1-flash"
curl -fsSL https://anyr.dev/cli.sh | bash
anyr login
anyr codex --model "deepseek/deepseek-v4.1-flash"

More: AI SDK · OpenAI SDK · coding agents · BYOK setup.

Bar chart of global KV cache bytes per token for DeepSeek-V1, V3.2, V4-Flash and V4.1-Flash, with arrows showing each generation is smaller
DeepSeek-V4.1-Flash KV cache size per token. Source: DeepSeek.

Related posts

Looking for the full catalog? Browse all models.