DeepSeek V4.1 Flash
DeepSeek🇨🇳 CNFrontier· Sep 10, 2026
TextReasoningVisionopen weightsAPI
Multimodal MoE model that natively processes images and text with a 1M-token context window. Based on a 552B backbone, it activates 8B/16B per step, cutting KV-cache overhead to roughly a quarter of its predecessor.
Summary of the model card
- Context
- 1M tokens
- Price / 1M tok
- $0.04 in · $0.29 out
- License
- mit
- Modalities
- text, image → text
3 variants
Sources
HuggingFaceOpenRouterDeutsch
Multimodales MoE-Modell, das Bilder und Text nativ verarbeitet und bis zu eine Million Tokens Kontext unterstützt. Es basiert auf einem 552B-Backbone, aktiviert pro Schritt nur 8B bzw. 16B Parameter und reduziert den KV-Cache-Aufwand auf etwa ein Viertel der Vorgängerversion.
Benchmarks
| Benchmark | Domain | Value | Evidence |
|---|---|---|---|
| LongBench v2vendor number | long-context | 44.7 % correct | 90 |
| SimpleQAvendor number | knowledge | 30.1 % correct | 81 |
| MMMU-Provendor number | vision | 56.5 % correct | 80 |
| DocVQAvendor number | vision | 95.6 ANLS | 71 |
| BigCodeBenchvendor number | coding | 56.8 % pass@1 | 66 |
| HumanEvalvendor number | coding | 69.5 % pass@1 | 57 |
| GPQA Diamondvendor number | reasoning | 93.4 % correct | 50 |
| MGSMvendor number | multilingual | 84.4 % correct | 44 |
| MMLUvendor number | knowledge | 68.3 % correct | 43 |
Vendor-reported numbers are marked and count with factor 0.7 in rankings. The evidence score says how much a benchmark still tells you today.
Provenance
- Detected
- Sep 27, 2026
- First source
- openrouter
- Description
- Summary of the model card
- Editorially reviewed
- —