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 V4.1 Flash's technical report — titled "Pushing Limits of KV Cache Compression" — describes how the new model compresses this cache far more aggressively than previous architectures. The result: faster inference speeds, lower memory requirements, and reduced cost per token — all without sacrificing output quality.
Performance: Surpassing V4 Pro on Every Metric
DeepSeek's official announcement states that V4.1 Flash "comprehensively surpassed V4 Pro across all key metrics, including performance, cost, speed, and task completion time." While specific benchmark numbers were not disclosed in the launch announcement, the company released a detailed technical report on Hugging Face for independent verification.
The model also features native multimodal support, meaning it can process both text and images natively rather than relying on separate encoder modules.
Pricing: How Much Does V4.1 Flash Cost?
DeepSeek has positioned V4.1 Flash as a cost leader in the frontier model market:
- Input (cache hit): $0.003 per unit
- Input (cache miss): $0.15 per unit
- Output: $0.60 per unit
- Peak hours (UTC 1“4 AM and 6“10 AM, Monday“Friday): Double the off-peak rates
For developers building on DeepSeek's API, the cache hit pricing is particularly notable — at $0.003 per unit, it is among the cheapest input costs of any frontier-class model available today.
V4 Pro Sunset: What Happens on September 14?
DeepSeek had originally planned to discontinue V4 Pro alongside the V4.1 Flash launch, but postponed the shutdown to September 14, 2026 at 12:00 Beijing Time. After that date:
- All API requests targeting V4 Pro will be automatically rerouted to V4.1 Flash
- Billing will switch to V4.1 Flash pricing (which is lower)
- No developer action is required — the transition is seamless
Where to Access V4.1 Flash
The model weights have been released on Hugging Face under the deepseek-ai/DeepSeek-V4.1-Flash repository. Developers can also access the model through Vercel's AI Gateway and DeepSeek's own API platform.
During the internal beta phase, DeepSeek limited the model to 20 concurrent requests per account — a constraint that may be relaxed as the model stabilizes in production.
Want to stay ahead of every new AI model release and pricing change? Follow AI Explainer India for timely, plain-English breakdowns of every major AI development that affects developers and businesses in India.