TL;DR
DeepSeek released V4.1 Flash on September 10, 2026. On DeepSeek's official benchmarks it beats V4 Pro across the board (DeepSWE v1.1: 74.2 vs 62.7), while its peak output price is about 30% of V4 Pro's. V4 Pro retires at 12:00 Beijing time on September 14, 2026, with requests auto-routed to V4.1 Flash at Flash rates.
Object overview
DeepSeek V4.1 Flash: Released September 10, 2026; new Causal Encoder-Decoder architecture; native vision understanding; open-sourced under MIT with weights on Hugging Face.
DeepSeek V4 Pro: Flagship of the V4 family, roughly 1.6T total parameters (~49B active), built for complex reasoning and agent tasks; service retires at 12:00 Beijing time on September 14, 2026.
DeepSeek V4 Flash (previous): V4 Flash general release from July 31, 2026, roughly 284B parameters (~13B active); now retired, with requests served by V4.1 Flash.
Core parameter comparison
| Dimension | V4 Pro | V4 Flash (old) | V4.1 Flash |
|---|---|---|---|
| Architecture | V4 MoE | V4 MoE | V4.1 Causal Encoder-Decoder (40 layers) |
| Total parameters | ~1.6T | ~284B | 552B (763B with vision encoder) |
| Active parameters | ~49B | ~13B | 8B input / 16B output |
| Context window | 1M tokens | 1M tokens | 1M tokens |
| Max output | 384K tokens | 384K tokens | 384K tokens |
| Vision | Not disclosed | Separate Vision-Exp build | Native (DeepSeek-ViT) |
| Open source | Not disclosed | Not disclosed | MIT, weights open (Hugging Face) |
| Concurrency limit | 500 | 2,500 | 2,500 |
| Output price (off-peak) | $1.98 / 1M tokens | $0.66 / 1M tokens | $0.60 / 1M tokens (CNY 4) |
| Output price (peak) | $3.96 / 1M tokens | $1.32 / 1M tokens | $1.20 / 1M tokens (CNY 8) |
| API status | Retires Sept 14, rerouted to V4.1 Flash | Retired, served by V4.1 Flash | GA, model id: deepseek-flash |
Key differences
- Outperforms across the board: On DeepSeek's official benchmarks, V4.1 Flash scores 74.2 on DeepSWE v1.1 vs 62.7 for V4 Pro and 54.4 for V4 Flash; 90.6 on Terminal-Bench 2.1 (V4 Pro: 87.9) and 88.1 on CyberGym (V4 Pro: 83.3).
- Much lower cost: Peak output at $1.20 per 1M tokens is about 30% of V4 Pro ($3.96); cache-hit input at $0.006 is about 86% lower than V4 Pro.
- New architecture: Asymmetric activation (8B input / 16B output) optimizes compute for agent workloads that read large context and generate limited output.
- Vision merged into the main model: Requests to old V4 Flash and its vision experimental build are all served by V4.1 Flash, unifying text and image processing in one primary API model.
Buying advice
- Developers needing cost-effective daily and agent tasks: use V4.1 Flash (model id: deepseek-flash) — stronger performance at a lower price.
- Production environments still calling V4 Pro: after 12:00 Beijing time on September 14, 2026, requests auto-reroute to V4.1 Flash and are billed at Flash rates — no code change needed, but validate outputs with the same prompts in advance.
- As of September 10, 2026, V4.1 Flash benchmarks are vendor-reported only; no independent third-party replication has been published.
Data sources & references
- 36Kr (Tencent Tech): DeepSeek V4.1 Flash release deep-dive
- Apidog: DeepSeek-V4.1-Flash vs V4-Pro vs V4-Flash
- 36Kr: DeepSeek cuts Flash pricing by up to 60%
- ZOL: DeepSeek releases V4.1 Flash