DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash, released on 10 September 2026 under MIT, is the smallest model in DeepSeek's new encoder-decoder MoE architecture family. It has 552B backbone parameters plus a 196B lookup memory, activates about 16B parameters per generated token, and accepts images. It now serves the deepseek-flash API model with a 1M-token context.
Specifications
- Context window
- 1Mabout 786,000 words, or 1,570 pages
- Max output
- 375Ktokens per response
- Input price
- $0.30per 1M tokens
- Output price
- $1.20per 1M tokens
- Cached input
- $0.006per 1M tokens
- Weights
- OpenMIT
- Parameters
- 552B total, 16B activeMoE
- Input types
- Text, Images
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Sep 10, 2026Current
DeepSeek API peak price (model deepseek-flash). Off-peak rates are half: $0.15 input, $0.003 cached, $0.60 output. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. 552B backbone excludes a 196B Engram memory table (763B in safetensors); 16B active at decode, 8B at prefill. Max output is DeepSeek's stated API limit of 384K.
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $32 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $9.12 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $25 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- Very low first-party API price, halved off-peak
- 1M-token context with a much smaller KV cache than V4-Flash
- MIT license with image input
Trade-offs to weigh
- 763B stored parameters make local hosting a multi-GPU job
- New architecture may need recent inference engine versions
Where it fits
DeepSeek positions DeepSeek-V4.1-Flash for:
Where you can use it
- Hugging Face
- Ollama
- API
- Together AI
- Fireworks
Run it yourself
Released under MIT. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 1,104 GB |
| 8-bit (FP8/INT8) | 552 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 335 GB |
This is a mixture-of-experts model: all 552B total parameters must sit in memory, but only about 16B are used per token, so it runs faster than a dense model of the same size.
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
| Model | Tier | Context | Input / output per 1M | |
|---|---|---|---|---|
| MiniMax-M3MiniMax | Advanced | 1M | $0.30 / $1.20 | Compare |
| Mistral Large 3Mistral AI | Advanced | 256K | $0.50 / $1.50 | Compare |
| NVIDIA Nemotron 3 UltraNVIDIA | Advanced | 1M | $0.50 / $2.20 | Compare |
| gpt-oss-120bOpenAI | Advanced | 128K | $0.15 / $0.60 | Compare |
Frequently asked questions
How much does DeepSeek-V4.1-Flash cost?
DeepSeek-V4.1-Flash lists at $0.30 per million input tokens and $1.20 per million output tokens, with cached input at $0.006 (checked October 10, 2026). DeepSeek API peak price (model deepseek-flash). Off-peak rates are half: $0.15 input, $0.003 cached, $0.60 output. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. 552B backbone excludes a 196B Engram memory table (763B in safetensors); 16B active at decode, 8B at prefill. Max output is DeepSeek's stated API limit of 384K. Confirm current rates on the provider's pricing page before you commit.
What is the context window of DeepSeek-V4.1-Flash?
DeepSeek-V4.1-Flash accepts up to 1,048,576 tokens of context (about 786,000 words, or 1,570 pages), and it can write up to 375K tokens in a single response. Long prompts cost more and can be slower, so most apps send only what the task needs.
Is DeepSeek-V4.1-Flash open source?
DeepSeek-V4.1-Flash is an open-weight model released under the MIT license. You can download the weights from Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run DeepSeek-V4.1-Flash locally?
Yes, if you have enough memory. The weights alone need about 335 GB at 4-bit quantization and about 1,104 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is DeepSeek-V4.1-Flash best for?
DeepSeek positions DeepSeek-V4.1-Flash for coding, agents, high volume, long documents. It accepts text, images as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- api-docs.deepseek.com/quick_start/pricing
- api-docs.deepseek.com/news/news260910
- api-docs.deepseek.com/updates
- huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/raw/main/config.json
- huggingface.co/api/models/deepseek-ai/DeepSeek-V4.1-Flash
- docs.together.ai/docs/serverless-models
- docs.fireworks.ai/serverless/pricing
- ollama.com/library
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.