DeepSeek-V4.1-Flash is a multimodal mixture-of-experts model from DeepSeek-AI that accepts text and images, generates text autoregressively, supports a combined context of up to one million tokens, and provides continuously controllable reasoning effort.
This model is ready for commercial or non-commercial use.
This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see link to Non-NVIDIA DeepSeek-V4.1-Flash Model Card
GOVERNING TERMS: [TO BE PROVIDED BY PRODUCT LEGAL VIA NSPECT WORKFLOW]
Global
Use Case: DeepSeek-V4.1-Flash is intended for long-context multimodal reasoning, coding, document understanding, and tool-using agentic applications that consume text and images and generate text.
Build.NVIDIA.com: 09/18/2026 via link
Huggingface: 09/10/2026 via link
References:
Architecture Type: Transformer
Network Architecture: Mixture-of-Experts
Total Parameters: 552B
Active Parameters: 8B (prefill), 16B (decode)
Vocabulary Size: 129,280
Input Types: Image, Text
Input Formats: Red, Green, Blue (RGB), String
Input Parameters: Two-Dimensional (2D), One-Dimensional (1D)
Other Input Properties: The vision configuration supports up to 1,024 image tokens and specifies a minimum image area of 295,936 pixels; the reference encoder supports interleaved image and text content.
Output Types: Text
Output Format: String
Output Parameters: One-Dimensional (1D)
Other Output Properties: Combined input and output context length is 1,048,576 tokens; the prompt protocol supports thinking content, tool calls, and structured API responses.
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Runtime Engines:
Supported Hardware:
Preferred Operating Systems: Linux
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
DeepSeek-V4.1-Flash v4.1
Data Modality: Image, Text
Image Training Data Size: Undisclosed
Text Training Data Size: Undisclosed
Training Data Collection: Undisclosed
Training Labeling: Undisclosed
Training Properties: The model was trained from scratch on a multimodal corpus comprising 45T tokens, with sparse attention trained at a 64K-token sequence length and context extended to 1M tokens at 34T tokens. Post-training used supervised fine-tuning, reinforcement learning, and on-policy distillation with large-scale automated synthesis of agent tasks and environments; the source does not disclose corpus names, licenses, or collection and labeling procedures.
Testing Data Collection: Undisclosed
Testing Labeling: Undisclosed
Testing Properties: Undisclosed
Evaluation Data Collection: Hybrid: Automated, Manually-Collected
Evaluation Labeling: Hybrid: Automated, Manually-Labeled
Evaluation Properties: Base models were evaluated in DeepSeek's internal framework under the same settings, and scores within 0.3 were considered equivalent. Instruct evaluations used reasoning_effort=100, temperature=1.0, and top_p=0.95; the benchmark results and methodology below are reported by DeepSeek-AI and do not represent independent NVIDIA verification.
Evaluation Benchmark Score: DeepSeek-V4.1-Flash-Base reported scores across knowledge, reasoning, coding, mathematics, long-context, and multimodal benchmarks; the instruct model reported 90.9 GPQA Diamond Pass@1, 90.6 Terminal-Bench 2.1 Pass@1, 74.2 DeepSWE v1.1 resolved, 88.1 CyberGym Pass@1, 54.8 AutomationBench Pass@1, and 31.8 Agent's Last Exam Pass@1.
| Benchmark (Metric) | Shots | DeepSeek-V4.1-Flash-Base |
|---|---|---|
| AGIEval (EM) | 3–5-shot | 83.4 |
| MMLU-Pro (EM) | 5-shot | 74.1 |
| C-Eval (EM) | 5-shot | 92.1 |
| MultiLoKo (LLM-Judge) | 5-shot | 45.5 |
| SimpleQA-Verified (EM) | 25-shot | 42.3 |
| SuperGPQA (EM) | 5-shot | 53.1 |
| BBH (EM) | 3-shot | 86.1 |
| BBEH (EM) | 1-shot | 27.2 |
| DROP (F1) | 1-shot | 87.9 |
| HellaSwag (EM) | 0-shot | 87.2 |
| BigCodeBench (Pass@1) | 3-shot | 60.6 |
| HumanEval (Pass@1) | 0-shot | 79.4 |
| GSM8K (EM) | 8-shot | 93.0 |
| MATH (EM) | 4-shot | 61.1 |
| MGSM (EM) | 8-shot | 80.2 |
| LongBench-V2 (EM) | 1-shot | 45.2 |
| MMMU-Pro (EM) | 4-shot | 56.5 |
| CVBench (EM) | 4-shot | 77.9 |
| DocVQA (LLM-Judge) | 4-shot | 95.6 |
| RefCOCO-avg (Acc@0.5) | 0-shot | 86.0 |
| Benchmark (Metric) | DeepSeek-V4.1-Flash |
|---|---|
| GPQA Diamond (Pass@1) | 90.9 |
| HLE (Pass@1) | 36.8 (39.1†) |
| Codeforces (Rating) | 3471 |
| MathArena Apex (Pass@1) | 65.6 |
| Terminal-Bench 2.1 (Pass@1) | 90.6 |
| Terminal-Bench 3.0 (Pass@1) | 30.0 |
| Terminal-Bench 4.0 (Pass@1) | 31.2 |
| DeepSWE v1.1 (Resolved) | 74.2 |
| ProgramBench (Almost@1) | 20.3 |
| NL2Repo-Bench (Score) | 64.0 |
| CyberGym (Pass@1) | 88.1 |
| SEC-Bench Pro (Pass@1) | 62.8 |
| ExploitGym (Pass@1) | 15.3 |
| HLE with tools (Pass@1) | 63.9 |
| AutomationBench (Pass@1) | 54.8 |
| Agent's Last Exam (Pass@1) | 31.8 |
| Chartography with tools (Pass@1) | 78.9 |
| BabyVision with tools (Pass@1) | 89.6 |
| ZeroBench-main with tools (Pass@5) | 49.0 |
† Text-only subset of HLE.
temperature=1.0, top_p=0.95, a 1M-token context limit, and max_steps=500. Terminal-Bench 2.1 ran without network access.| Benchmark (Metric) | Claude Code | Codex | OpenCode | Pi | mini-SWE | DSH Minimal | DSH Standard | DSH PTC |
|---|---|---|---|---|---|---|---|---|
| DeepSWE v1.1 (Resolved) | 69.8 | 65.6 | 65.5 | 66.2 | 74.2 | 72.6 | 70.5 | 67.6 |
| Terminal-Bench 2.1 (Pass@1) | 88.0 | 84.1 | 85.0 | 86.1 | 90.3 | 90.6 | 85.8 | 85.8 |
Acceleration Engine: vLLM
Test Hardware: NVIDIA Hopper (H200)
DeepSeek-V4.1-Flash uses a 40-layer Causal Encoder-Decoder Transformer with a 20-layer causal encoder and a 20-layer decoder. It combines Compressed Sparse Attention 2, SWA Bounded Replay, FP4 main KV caching, Single-Pass mHC, a 196B-parameter Engram conditional memory, and DSpark speculative decoding; the source reports a persistent global KV cache footprint of 890 bytes per token.
Each mixture-of-experts layer has one shared expert and 384 routed experts, with six routed experts activated per token. DeepSeek-ViT, trained from scratch with two-dimensional rotary position embeddings and 3×3 pixel-unshuffle downsampling, feeds visual embeddings through a two-layer MLP projector for joint image-and-text processing.
This release does not include a Jinja-format chat template. Its reference encoding supports multi-turn conversations, tool calling, thinking mode, numeric reasoning effort from 1 to 100, mid-conversation system messages, and interleaved image content; model inference, tool execution, and HTTP transport remain the caller's responsibility.
The source recommends temperature=1.0, top_p=0.95 or 1.0, a 1M-token context window, and max_tokens of at least 256K.
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please make sure you have proper rights and permissions for all input image content; if image includes people, personal health information, or intellectual property, the image generated will not blur or maintain proportions of image subjects included.
Users are responsible for model inputs and outputs. Users are responsible for ensuring safe integration of this model, including implementing guardrails as well as other safety mechanisms, prior to deployment.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here