DeepSeek V4.1 Flash
DeepSeek's multimodal Flash model with a 1M context, MIT weights and very low API prices
Context window
1M tokens
Input / 1M tokens
$0.30
Output / 1M tokens
$1.20
Provider
DeepSeek
DeepSeek API price for deepseek-flash at peak hours. Off-peak rates are half ($0.15/$0.60), and cache-hit input costs $0.006 per 1M at peak ($0.003 off-peak). The October 17 update (V4.1-Flash-1017) kept the same price. The MIT open weights are free to self-host. · Data verified 2026-11-08
DeepSeek V4.1 Flash (September 10, 2026) is the smallest model in DeepSeek's new architecture family and its first Flash model with native image understanding. It replaced V4 Flash on the DeepSeek API, where it is served as deepseek-flash, and DeepSeek says tests by multiple parties put it ahead of V4 Pro on performance, cost, speed and total runtime. The checkpoint has 763B parameters around a 552B Mixture-of-Experts backbone, but only 8B are active per token for prefill and 16B for decode, with a 1M-token context. An October 17 update, V4.1-Flash-1017, raised DeepSWE v1.1 from 74.2% to 77.0% and Terminal-Bench 4.0 from 31.2% to 42.3% at the same price. The weights are on Hugging Face under the MIT license.
How to access it
Call the DeepSeek API with model deepseek-flash (the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route here), chat with it at chat.deepseek.com, or download the MIT weights from Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and serve them with vLLM or SGLang. Also on OpenRouter (deepseek/deepseek-v4.1-flash-1017) and Alibaba QwenCloud.
Strengths
- ✓Native image input, new for DeepSeek Flash models
- ✓77.0% on DeepSWE v1.1 and 42.3% on Terminal-Bench 4.0 after the 1017 update (DeepSeek-reported)
- ✓Low API price ($0.30/$1.20 per 1M at peak, half off-peak) with cache hits at $0.006
- ✓Small active footprint (8B prefill, 16B decode) for fast, cheap serving
- ✓MIT-licensed open weights
Best for developers who...
When to choose it (and when not to)
Reach for DeepSeek V4.1 Flash when...
- →You want strong agentic coding at a very low per-token price
- →You need image input from a low-cost open model
- →You process long inputs up to 1M tokens
- →You want MIT-licensed weights you can self-host or fine-tune
Look elsewhere if...
- ✕You need the broadest world knowledge: V4 Pro scores higher on DeepSeek's base-model knowledge tests such as SimpleQA-Verified
- ✕You want to self-host but lack multi-GPU hardware for a 763B-parameter checkpoint
- ✕You need audio input
- ✕Most of your traffic falls in DeepSeek's peak hours, when prices are twice the off-peak rate
How to use it
- ›DeepSeek recommends temperature 1.0 and top_p 1.0
- ›Allow a large max_tokens budget for hard reasoning tasks (DeepSeek suggests 256K or more)
- ›Run batch jobs in off-peak hours to pay half price
- ›Keep system prompts and long prefixes stable to get cache-hit pricing
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI(
api_key="YOUR_DEEPSEEK_API_KEY",
base_url="https://api.deepseek.com",
)
resp = client.chat.completions.create(
model="deepseek-flash",
messages=[{"role": "user", "content": "Find the bug in this function and fix it."}],
)
print(resp.choices[0].message.content)OpenAI-compatible endpoint. For self-hosting: vllm serve deepseek-ai/DeepSeek-V4.1-Flash-1017.
API model id: deepseek-flash
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| DeepSWE v1.1 | 77.0% (1017 update) | DeepSeek-reported, max effort; 74.2% for the September release and 54.4% for V4 Flash |
| Terminal-Bench 4.0 | 42.3% (1017 update) | DeepSeek-reported; 31.2% for the September release |
| HLE (no tools) | 42.6% (1017 update) | DeepSeek-reported; 36.8% for the September release |
| GPQA Diamond | 90.9% | DeepSeek-reported for the September release, max effort |
| CyberGym | 88.1% | DeepSeek-reported for the September release |
| BabyVision (with tools) | 92.5% (1017 update) | DeepSeek-reported; 89.6% for the September release |
Tools powered by DeepSeek V4.1 Flash
Compare DeepSeek V4.1 Flash
Compare DeepSeek V4.1 Flash with any other model
Build a comparison →All model comparisons →