DeepSeek V4.1 Flash: The 552B Model That Replaced V4-Pro — Specs, Pricing, and How to Run It
DeepSeek V4.1 Flash: The 552B Model That Replaced V4-Pro — Specs, Pricing, and How to Run It
On September 10, 2026, DeepSeek released DeepSeek V4.1 Flash, the smallest model in its new architecture family. Four days later, the company started routing all DeepSeek V4 Pro traffic to it, and the previous V4 Flash and V4 Flash Vision Exp models were retired. In other words, V4.1 Flash is not just another checkpoint: it became the default model for most API users, with higher agentic scores and a noticeably lower cost per task.
Below are the real numbers from DeepSeek's official documentation, what changed in the lineup, the full pricing breakdown, and exactly how to run the model today.
Quick answer: DeepSeek V4.1 Flash is a mixture-of-experts model with a 552B parameter backbone, but only 8B parameters are active per token during prefill and 16B during decode. It supports a one-million-token context window, handles text and images natively, and ships under an MIT license. Weights are free on Hugging Face; API usage costs $0.30 per million input tokens and $1.20 per million output tokens at peak, with off-peak rates at half price.
What changed in DeepSeek's model lineup
The "Flash" tier in DeepSeek's new family is the lightest model, but lighter here does not mean weaker. A new Causal Encoder-Decoder architecture lets the model hit higher capability at a lower serving cost, and DeepSeek migrated users onto it in stages:
- V4 Flash and V4 Flash Vision Exp are retired. For compatibility, the old model names temporarily route to V4.1 Flash.
- Starting at 04:00 UTC on September 14, 2026, every V4 Pro request routes to V4.1 Flash at V4.1 Flash rates while V4 Pro is phased out.
- The API model name is now deepseek-flash.
Core specifications
| Spec | Value |
|---|---|
| Backbone parameters | 552B (mixture of experts) |
| Activated parameters | 8B during prefill / 16B during decode |
| Architecture | Causal Encoder-Decoder (20 encoder + 20 decoder layers) |
| Context window | Up to 1,048,576 tokens |
| Inputs | Text and images (natively multimodal) |
| Outputs | Text |
| Reasoning effort | Continuously controllable, 1 to 100 |
| KV cache size | 890 bytes per token |
| License | MIT (open weights) |
| API model name | deepseek-flash |
Why the KV cache is the real story
The expensive part of an agentic workload is not the question, it is the context. As a conversation grows, the model needs a larger KV cache, and that is what makes long agent runs slow and costly. V4.1 Flash compresses the global KV cache to 890 bytes per token, roughly a quarter of V4 Flash's footprint and about 437 times smaller than the first-generation DeepSeek-V1.
That compression comes from the Causal Encoder-Decoder design plus Compressed Sparse Attention with FP4 KV caching. The practical effect is simple: agents that read large codebases or long documents get cheaper to run, and a cache hit costs $0.006 per million input tokens.
Benchmarks against bigger models
All figures below come from DeepSeek's published tables at maximum reasoning effort. One benchmark never tells the whole story, but the direction is clear: a 552B model is beating far larger competitors on agentic tasks.
| Benchmark | V4.1 Flash | V4-Pro | Claude Opus-5.0 | GPT-5.6 Sol |
|---|---|---|---|---|
| Codeforces (rating) | 3471 | 3348 | — | — |
| DeepSWE v1.1 | 74.2 | 62.7 | 74.0 | 73.0 |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 89.1 | 88.8 |
| AutomationBench | 54.8 | 43.2 | 50.3 | 45.8 |
| CyberGym | 88.1 | 83.3 | — | 84.5 |
| Agent's Last Exam | 31.8 | 25.7 | 28.6 | 26.7 |
| GPQA Diamond | 90.9 | 92.4 | 93.4 | 94.1 |
The pattern is consistent: V4.1 Flash wins on long execution tasks such as DeepSWE, Terminal-Bench, and AutomationBench, while closed frontier models still lead on knowledge-heavy benchmarks like GPQA Diamond and Humanity's Last Exam. The question is not who is strongest, it is who is cheapest for the job you actually run.
Pricing per million tokens
| Token type | Peak | Off-peak |
|---|---|---|
| Uncached input | $0.30 | $0.15 |
| Cached input | $0.006 | $0.003 |
| Output | $1.20 | $0.60 |
Off-peak rates are exactly half of peak rates, so batch jobs, document analysis, and code indexing are worth scheduling off-peak. Third-party hosts also set their own rates: DeepInfra lists $0.20 input and $0.60 output on its standard tier, and Baseten charges $0.30 input and $1.20 output, so compare before you commit.
Note: AI API pricing changes fast. The numbers above are from DeepSeek's official release and pricing pages, so check the current official pricing page before you build a cost model on them.
How to run it: three practical options
1) Official API:
Set the model name to deepseek-flash and keep everything else the same. DeepSeek's API docs cover the exact call format.
2) Ollama:
The model is available on Ollama cloud as deepseek-v4.1-flash:cloud, with a second tag deepseek-v4.1-flash:1017-cloud, and it plugs straight into coding agents.
3) Self-host the weights:
The MIT-licensed weights are on Hugging Face, but they are 552B parameters, so plan for data-center hardware, not a laptop.
ollama launch opencode --model deepseek-v4.1-flash:cloud
ollama launch claude --model deepseek-v4.1-flash:cloud
If you prefer the hosted endpoint, the request format is the same as any other Ollama chat call:
-H "Authorization: Bearer $OLLAMA_API_KEY" \
-d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"Hello!"}]}'
Does it see images?
Yes. Vision is not bolted on after training: the model was trained from scratch on a multimodal corpus of 45 trillion tokens. DeepSeek's published base-model numbers include 95.6 on DocVQA and 86.0 on RefCOCO-avg, which means you can drop a document scan or a chart into the conversation and ask about it directly.
Things to watch out for
Warning:
- Running it locally is not realistic for most people. At 552B parameters, plan on a hosted endpoint unless you already rent serious GPU capacity.
- Any host you use can see your prompts. If you send customer data or internal documents, read the provider's privacy policy and retention terms first.
- Closed models still lead on knowledge-heavy benchmarks. If your workload is fact-heavy Q&A, evaluate both before switching.
- The legacy name deepseek-v4-flash only routes to the new model temporarily, so write deepseek-flash in your code today.
FAQ
Is DeepSeek V4.1 Flash free?
Answer: The weights are free and open-source under the MIT license on Hugging Face, so you can download and run them without license fees. API access is paid but cheap: $0.30 per million input tokens and $1.20 per million output tokens at peak.
How is it different from DeepSeek V4 Pro?
Answer: V4 Pro has a 1.6T backbone, yet V4.1 Flash beats it on most agentic benchmarks such as Terminal-Bench, DeepSWE, and AutomationBench while costing less. Since September 14, 2026, DeepSeek routes V4 Pro requests to V4.1 Flash as V4 Pro is phased out.
Can I run it on my own laptop?
Answer: Realistically, no. A 552B-parameter model needs far more memory and compute than consumer hardware provides. Use the API or a hosted provider such as Ollama cloud or DeepInfra.
Does it work with coding agents like Claude Code and OpenCode?
Answer: Yes. DeepSeek announced official support from partners including WorkBuddy (and its CodeBuddy product) and OpenCode, and on Ollama you can launch either agent with a single command plus the model name.
What is the 1017 tag on Ollama?
Answer: Ollama lists two versions of the model: deepseek-v4.1-flash:cloud as the main tag, and deepseek-v4.1-flash:1017-cloud as a separate update point. Both list a one-million-token context window and text plus image input. No detailed official changelog comparing the two has been published, so check DeepSeek's official change log for updates.
Bottom line
DeepSeek V4.1 Flash is not simply a smaller model. It is a 552B mixture-of-experts system that activates 8B parameters on input, handles a million tokens of context, understands images natively, and cuts KV cache to 890 bytes per token, which is what pushed agentic costs down. If you already call the DeepSeek API, update the model name to deepseek-flash, set the reasoning effort to match the task instead of maxing it out by default, and schedule heavy jobs off-peak. If you use coding agents, the fastest path is a single Ollama command rather than wiring up a new account from scratch.
البيت | ايموني
https://www.aymany.com/
