Chinese AI lab DeepSeek has released V4.1 Flash, its latest large language model that the company says surpasses the previous V4 Pro across every major benchmark while costing significantly less to run. Launched on September 10, 2026, the model introduces a new architecture built around KV cache compression, a technique that dramatically reduces the memory and compute required for inference.
What Is KV Cache Compression and Why Does It Matter?
When an AI model processes a conversation, it stores previous tokens in a "Key-Value cache" so it doesn't have to reprocess the entire conversation from scratch each time. The problem is that this cache grows linearly with conversation length, consuming expensive GPU memory.
DeepSeek's API announcement says V4.1 Flash's cache needs about one quarter of the HBM (GPU memory) and one eighth of the SSD storage of the previous generation; Techstrong.ai put it at roughly 890 bytes per token. Less memory per conversation means more users per GPU, cheaper long contexts and lower cost per token, which is what lets DeepSeek cut prices.
Performance: DeepSeek's Claims Against V4 Pro
DeepSeek's announcement says V4.1 Flash beats V4 Pro on performance, cost, speed and task completion time, citing third-party tests. These are the company's own claims; we have not seen independent benchmarks that confirm them across the board, so test it on your own workload.
The model also features native multimodal support, so it reads images as well as text. Under the hood it is a 552-billion-parameter mixture-of-experts model that DeepSeek describes as a causal encoder–decoder, activating about 8 billion parameters to read input and 16 billion to write output, which is why it can be both large and cheap to run.
Pricing: How Much Does V4.1 Flash Cost?
DeepSeek cut prices with this release and kept its peak and off-peak structure: off-peak usage costs half the peak rate, and cached input is far cheaper than fresh input. The Next Web covered the cut. Because the per-million-token figures have changed more than once, check DeepSeek's API pricing page for current numbers before you budget, and schedule batch jobs for off-peak hours where you can.
V4 Pro Sunset: What Happens on September 14?
From 04:00 UTC (12:00 Beijing time) on September 14, 2026, DeepSeek changed what its model names point to:
- Requests to
deepseek-v4-prorun on V4.1 Flash and are billed at V4.1 Flash rates, until DeepSeek launches V4.1 Pro - The older V4 Flash and V4 Flash Vision models are retired and route to V4.1 Flash
- Your code keeps working without changes, but the model behind the name is different
Our take: a checklist if your app is pinned to V4 Pro
“No action required” is true for uptime, not for quality. As Requesty pointed out, a pinned model ID that silently points at a different model can change your outputs overnight. Here is what we would do:
- Rerun your evals. Compare answers from before and after September 14 on your own test set, especially for long reasoning tasks where a Pro model used to have the edge.
- Check your bills. V4.1 Flash rates are lower, so a cost drop is expected; a cost rise means longer outputs.
- Name the model explicitly. Switch to
deepseek-flashso your logs show what actually ran. - Plan for V4.1 Pro. When it launches, the
deepseek-v4-proname may move again; keep a test ready. - Move batch work off-peak to halve its cost.
Where to Access V4.1 Flash
The model weights are on Hugging Face in the deepseek-ai/DeepSeek-V4.1-Flash repository; check the model card for the licence before using them commercially. Through DeepSeek's own API, call it with the model name deepseek-flash. Many third-party routers also list it, though availability varies.
Sources
- DeepSeek API news: DeepSeek V4.1 Flash release (September 10, 2026)
- Techstrong.ai: DeepSeek unveils V4.1 Flash with architectural upgrades and price cuts
- The Next Web: DeepSeek V4.1 Flash launches, V4 Pro retired, prices cut
- Requesty: DeepSeek V4.1 Flash retires V4 Pro behind pinned model IDs
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash
Frequently asked questions
Do I need to change my code to use V4.1 Flash?
Not to keep it running: the old V4 Flash and V4 Pro model names now route to V4.1 Flash. But the model behind those names has changed, so rerun your tests, and switch to the deepseek-flash model name if you want your logs to show what actually ran.
What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is the latest model from Chinese lab DeepSeek, released on September 10, 2026. It is a 552-billion-parameter mixture-of-experts model with native image understanding, and its redesigned KV cache needs about a quarter of the GPU memory of the previous generation.
How much does DeepSeek V4.1 Flash cost?
DeepSeek cut prices with this release and uses peak and off-peak rates, with off-peak at half the peak price. Check DeepSeek's API pricing page for current per-million-token figures before budgeting, since they have changed more than once.
Is DeepSeek V4.1 Flash better than V4 Pro?
DeepSeek says it beats V4 Pro on performance, cost, speed and task completion time, citing third-party tests. Independent benchmarks will tell the fuller story. V4 Pro requests have run on V4.1 Flash since September 14, 2026.
What is KV cache compression in DeepSeek V4.1 Flash?
The KV (key-value) cache stores the context a model has already processed so it does not recompute it. DeepSeek says V4.1 Flash's cache needs about 1/4 of the GPU memory (HBM) and 1/8 of the SSD storage of the previous generation, which makes long conversations cheaper to serve.
When will DeepSeek V4 Pro be discontinued?
From 04:00 UTC (12:00 Beijing time) on September 14, 2026, requests to deepseek-v4-pro run on V4.1 Flash and are billed at V4.1 Flash rates, until DeepSeek launches V4.1 Pro.