DeepSeek Announced Sep 10, 2026
DeepSeek V4.1 Flash
Available on MiniRouter for text generation. DeepSeek reports stronger coding results than V4 Flash.
Routable on MiniRouter
Use deepseek/deepseek-v4.1-flash via the API →- Architecture
- 552B MoE
- Causal Encoder–Decoder; 8B active input parameters, 16B output.
- Model limits
- 1M context / 384K output
- Text and image input. MiniRouter supports text only.
- DeepSeek API ID
- deepseek-flash
- MiniRouter: deepseek/deepseek-v4.1-flash.
- Weights
- MIT license
- Open weights and technical report.
Publisher pricing
| Phase | Input / 1M | Output / 1M | Blended / 1M | Effective | Source |
|---|---|---|---|---|---|
| Off-peak | $0.15 | $0.60 | $0.26 | Sep 10, 2026 – open-ended | DeepSeek ↗ |
| Peak · weekdays 01–04 / 06–10 UTC | $0.30 | $1.20 | $0.52 | Sep 10, 2026 – open-ended | DeepSeek ↗ |
Publisher-reported benchmarks
| Benchmark | Reported result | Reported by | As of | Source |
|---|---|---|---|---|
| DeepSWE v1.1 | 74.2% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
| Terminal Bench 2.1 | 90.6% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
| Terminal Bench 3.0 | 30.0% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
| Terminal Bench 4.0 | 31.2% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
| GPQA Diamond | 90.9% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
| CyberGym | 88.1% | DeepSeek | Sep 10, 2026 | Announcement ↗ |
- — DeepSeek's API replaces V4 Flash and V4 Flash Vision Exp with V4.1 Flash. The deepseek-v4-pro alias switches on September 14 at 04:00 UTC.
- — Prices apply from September 10 at 04:00 UTC. Cached input: $0.003/M off-peak, $0.006/M peak. MiniRouter adds 5%.
- — GPQA Diamond: 90.9%, versus V4 Pro's 92.4% in DeepSeek's results.
Related coverage
Sources
- Read at DeepSeek ↗
DeepSeek V4.1 Flash model card
Published Sep 10, 2026
- Read at DeepSeek ↗
DeepSeek Models & Pricing
Published Sep 10, 2026