Model Card: DeepSeek-V4.1-Flash

TL;DRDeepSeek-V4.1-Flash (`deepseek-flash`) is a 552B MoE with 8B prefill and 16B decode active parameters, a 1M-token context window, and 384K maximum output. It is MIT-licensed, priced at $0.30 input and $1.20 output per million peak-hour tokens, and brings CED, CSA2, and FP4 KV-cache compression to agent workloads.

#ai-agent #model-release #daily #deepseek #model-family-deepseek

Table of Contents

Model information
Highlights
Benchmark context
What it means for agent builders
Comparison with the prior model
Takeaway
References

Series: AI Model Tracker (19 / 24)

Model Card｜MiniCPM5-2B Model Card: Nex-N2.5-mini

🌏 中文版

Model information

Item	Value
Model ID	`deepseek-flash`; the Hugging Face repository is `deepseek-ai/DeepSeek-V4.1-Flash`
Parameters	552B total, 8B active for prefill and 16B for decode; 763B including the vision encoder
Context window	1,000,000 tokens; 384K maximum output
Price	Input: $0.30 peak / $0.15 off-peak per 1M tokens; output: $1.20 peak / $0.60 off-peak
License	MIT; weights and reference inference code are published on Hugging Face
Released	2026-09-10

Highlights

DeepSeek reports 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1. Its reported Claude Opus 5 comparisons are 89.1 and 74.0 respectively.
Causal Encoder-Decoder, Compressed Sparse Attention 2, and an FP4 primary KV cache reduce global KV-cache use to 890 bytes per token.
The API exposes a continuous reasoning_effort scale from 1 to 100, letting callers trade cost against reasoning depth without changing models.

Benchmark context

Benchmark	DeepSeek-V4.1-Flash	Prior V4-Flash-0731	Reported leading competitor
Terminal-Bench 2.1	90.6	82.7	Claude Opus 5: 89.1
DeepSWE v1.1	74.2	54.4	Claude Opus 5: 74.0
AutomationBench v1.0.6	54.8	37.7	Claude Opus 5: 50.3
CyberGym	88.1	76.7	GPT-5.6 Sol / GLM-5.3: 84.5
GPQA Diamond	90.9	89.9	GPT-5.6 Sol: 94.1

These are DeepSeek's release-day results, run with its own harness and high reasoning settings. They are useful for choosing candidates to evaluate, not independent reproductions.

What it means for agent builders

CED changes the long-context cost equation for multi-turn, tool-using sessions: the decoder projects from the encoder's final states instead of independently computing KV cache at every layer. A 1M-token context and 384K output limit make the model worth testing for repository-scale coding or long automation runs.

The reported agentic results make it a reasonable candidate for coding and CLI agents, particularly workloads currently routed to V4-Pro. DeepSeek says requests to deepseek-v4-pro will route to V4.1-Flash from September 14 at Flash pricing. Do not infer broad reasoning leadership from those results: its published GPQA Diamond score trails GPT-5.6 Sol, and its Humanity's Last Exam result trails Opus 5.

Comparison with the prior model

The reported gains concentrate on agentic tasks while using fewer active parameters than V4-Flash-0731. The architectural claim to test in practice is not only the benchmark score, but whether KV-cache compression improves the latency and cost profile of the actual long-context workload.

Takeaway

Treat V4.1-Flash as a strong agent-workload candidate with an unusually aggressive price change, then validate it against the task, tool loop, and reliability requirements that matter to your system.