DeepSeek-V4.1-Flash Released: 552B MoE Multimodal Model
At a Glance
| Item | Value |
|---|---|
| Repository | deepseek-ai/DeepSeek-V4.1-Flash |
| Published | 2026-09-10 |
| License | mit |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
deepseek-ai has released “DeepSeek-V4.1-Flash", a multimodal Mixture-of-Experts (MoE) model with a 552B backbone parameter count, on Hugging Face. It supports a context length of up to 1 million tokens, natively processes images and text, and generates text autoregressively.
Specifications
- Parameter Count: 552B backbone (active parameters: 8B during prefill / 16B during decode)
- Architecture: MoE (Causal Encoder-Decoder, 1 shared expert, 384 routed experts, 6 routed experts activated per layer)
- Context Length: Up to 1 million tokens
- License: mit
Performance
Comparison tables with base models and frontier models as presented in the model card and technical report are shown below.
→ Scroll horizontally to see all columns
| Benchmark (Metric) | # Shots | DeepSeek-V4-Flash-Base | DeepSeek-V4-Pro-Base | DeepSeek-V4.1-Flash-Base |
|---|---|---|---|---|
| Architecture | — | MoE | MoE | MoE |
| # Backbone Params | — | 284B | 1.6T | 552B |
| # Activated Params | — | 13B | 49B | 8B / 16B |
| World Knowledge | ||||
| AGIEval (EM) | 3–5-shot | 83.9 | 84.4 | 83.4 |
| MMLU-Pro (EM) | 5-shot | 68.3 | 73.5 | 74.1 |
| C-Eval (EM) | 5-shot | 92.1 | 93.1 | 92.1 |
| MultiLoKo (LLM-Judge) | 5-shot | 42.6 | 50.9 | 45.5 |
| SimpleQA-Verified (EM) | 25-shot | 30.1 | 55.2 | 42.3 |
| SuperGPQA (EM) | 5-shot | 46.5 | 53.9 | 53.1 |
| Language & Reasoning | ||||
| BBH (EM) | 3-shot | 86.9 | 87.5 | 86.1 |
| BBEH (EM) | 1-shot | 25.4 | 29.8 | 27.2 |
| DROP (F1) | 1-shot | 88.6 | 88.7 | 87.9 |
| HellaSwag (EM) | 0-shot | 85.7 | 88.0 | 87.2 |
| Code & Math | ||||
| BigCodeBench (Pass@1) | 3-shot | 56.8 | 59.2 | 60.6 |
| HumanEval (Pass@1) | 0-shot | 69.5 | 76.8 | 79.4 |
| GSM8K (EM) | 8-shot | 90.8 | 92.6 | 93.0 |
| MATH (EM) | 4-shot | 57.4 | 64.5 | 61.1 |
| MGSM (EM) | 8-shot | 85.7 | 84.4 | 80.2 |
| Long Context | ||||
| LongBench-V2 (EM) | 1-shot | 44.7 | 51.5 | 45.2 |
| Multimodal | ||||
| 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 |
Next, a restricted comparison table with frontier models (at maximum inference effort settings) is shown below.
→ Scroll horizontally to see all columns
| Benchmark (Metric) | GPT-5.6 Sol | K3 | DS-V4-Pro | DS-V4-Flash | DeepSeek-V4.1-Flash |
|---|---|---|---|---|---|
| Reasoning | |||||
| GPQA Diamond (Pass@1) | 94.1 | 92.9 | 92.4 | 89.9 | 90.9 |
| Codeforces (Rating) | — | — | 3348 | 3289 | 3471 |
| MathArena Apex (Pass@1) | — | 65.6 | 65.3 | 58.6 | 65.6 |
| Agentic | |||||
| Terminal-Bench 2.1 (Pass@1) | 88.8 | 88.3 | 87.9 | 82.7 | 90.6 |
| Terminal-Bench 3.0 (Pass@1) | 34.4 | 17.7 | 11.8 | 7.6 | 30.0 |
| DeepSWE v1.1 (Resolved) | 73.0 | 67.5 | 62.7 | 54.4 | 74.2 |
| CyberGym (Pass@1) | 84.5 | 80.0 | 83.3 | 76.7 | 88.1 |
| AutomationBench (Pass@1) | 45.8 | 46.7 | 43.2 | 37.7 | 54.8 |
| Agent’s Last Exam (Pass@1) | 26.7 | 27.6 | 25.7 | 25.2 | 31.8 |
According to the evaluation results published by the authors, DeepSeek-V4.1-Flash achieves the highest scores in its scale category and among comparisons in coding fields (HumanEval, BigCodeBench), competitive programming (Codeforces), and agentic tasks (such as Terminal-Bench and DeepSWE v1.1). While it also demonstrates high scores in general knowledge and reasoning like MMLU-Pro, it falls behind DeepSeek-V4-Flash and DeepSeek-V4-Pro in some items such as long-context tasks (LongBench-V2) and multilingual math (MGSM).
Strengths and Use Cases
- Multimodal Support: Multimodal tasks that process images and text simultaneously (DocVQA, RefCOCO, etc.)
- Coding and Agent Tasks: Extremely high performance in Python tasks involving multiple library calls, as well as terminal operation and software engineering agents
- Long-Context Processing: Supports context lengths up to 1 million tokens, improving cost efficiency through KV cache compression technologies (such as CSA2 and FP4 main KV cache)
- Inference Effort Control: Supports a “continuously controllable reasoning effort" setting that allows adjusting inference cost and accuracy using an integer from 1 to 100
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 763.2B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 1083GB of VRAM (multi-GPU or CPU offload required) | BF16 | 902.8GB | 1083.3GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-09-18): llama.cpp: not registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
How to Get It
- Distribution Format: safetensors / transformers
- How to Get: Available for download from the Hugging Face repository (
deepseek-ai/DeepSeek-V4.1-Flash). No license agreement is required for access (gated: false).
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-09-17 | nvidia | NVFP4 | nvidia/DeepSeek-V4.1-Flash-NVFP4 | BF16 1690.8GB (does not fit a single consumer GPU) |
In addition, 44 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.