{
  "id": "deepseek-4-1-flash",
  "title": "DeepSeek 4.1 Flash",
  "lang": "en",
  "source_lang": "en",
  "summary": "DeepSeek 4.1 Flash (DeepSeek-V4.1-Flash) is a multimodal Mixture-of-Experts language model released in September 2026, with a 552-billion-parameter backbone and a one-million-token context window.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "product",
  "aliases": [
    "DeepSeek-V4.1-Flash",
    "DeepSeek V4.1 Flash",
    "deepseek-flash (API model name)"
  ],
  "generated_at": "2026-10-16",
  "model": "pro/claude-haiku-5-5",
  "url": "https://aiwiki.page/en/deepseek-4-1-flash/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/deepseek-4-1-flash/",
      "json": "https://aiwiki.page/zh/deepseek-4-1-flash.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "DeepSeek 4.1 Flash, officially named DeepSeek-V4.1-Flash, is a large language model from the Chinese AI developer DeepSeek. It is built on a sparse Mixture-of-Experts design, accepts text and images, and supports a context window of about one million tokens. It belongs to the DeepSeek V4 family and is aimed at coding, reasoning, and long-running agent workloads."
    },
    {
      "heading": "Release and naming",
      "text": "The company's API change log records the launch on 10 September 2026 and describes the model as the smallest member of its new architecture family. The new architecture is designed for a higher capability ceiling, faster inference, higher throughput, and scaling to larger models. Developers select the model by changing the API name to deepseek-flash. The earlier V4 Flash and V4 Flash Vision Exp models have been retired; for compatibility, the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash.\n\nThe \"smallest\" label is relative to the family's flagship. The 552B backbone is far larger than the 284B total parameters of the previous V4 Flash line, so the label should not be read as a claim about absolute size."
    },
    {
      "heading": "Architecture",
      "text": "The model's backbone has 552 billion parameters, but only a fraction is used for any given token. Its 552B-parameter backbone activates 8B parameters per token during prefill and 16B during decoding. Taking the decode figure, roughly \\(\\frac{16}{552} \\approx 2.9\\%\\) of the backbone is active per generated token.\n\nThe network is described as a Causal Encoder–Decoder Transformer. It uses a 40-layer design with a 20-layer causal encoder and a 20-layer decoder. The architecture is asymmetric: it is designed to spend less compute on long inputs than on generated output, which suits prompts that are much longer than their responses. Each mixture-of-experts layer contains one shared expert and 384 routed experts, with six routed experts activated per token.\n\nFor attention, the first two encoder layers use sliding window attention (SWA); the rest use Compressed Sparse Attention 2 (CSA2). The model also combines a hierarchical sparse indexer with the CSA2 mechanism to reduce the cost of long-context inference. Other reported components include a 196B-parameter Engram conditional memory, and DSpark speculative decoding."
    },
    {
      "heading": "Key-value cache efficiency",
      "text": "The architecture's main stated goal is reducing inference cost for agent workloads, which are input-heavy. The authors note that prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. To address this, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. The resulting global cache footprint is 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. A separate deployment technique reduces the persistent cache kept on SSD or in host memory to roughly one-eighth of the predecessor's figure."
    },
    {
      "heading": "Multimodal capabilities",
      "text": "Unlike models that add vision after the main training, this model was trained with images from the start. The model features DeepSeek-ViT, a custom-trained vision encoder that processes visual embeddings jointly with text from the start of pre-training. The model accepts text and image inputs and produces text output. The vision configuration supports up to 1,024 image tokens and specifies a minimum image area of 295,936 pixels."
    },
    {
      "heading": "Context window and reasoning control",
      "text": "The model supports a context window of 1,048,576 tokens. DeepInfra serves the model at fp8 with the full 1,048,576-token context, so the served window matches the model's native window. Reasoning depth is adjustable. DeepSeek-V4.1-Flash supports reasoning effort from 1 to 100. Developers can therefore trade speed and cost against accuracy per request. The predecessor V4 Flash 0731 release used a coarser scheme with three reasoning-effort levels: low, high, and max."
    },
    {
      "heading": "Training",
      "text": "The model was pretrained on a large multimodal corpus. The team pretrained DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conducted comprehensive post-training. Context lengths were extended during training to one million tokens, which DeepSeek describes as a design choice for input-heavy agentic work."
    },
    {
      "heading": "Performance",
      "text": "Most benchmark figures come from DeepSeek itself. Its change log lists, among other results, GPQA Diamond: 90.9, Codeforces (Rating): 3471, and Terminal-Bench 2.1: 90.6. DeepSeek positions the model as the cost-efficient tier of the V4.1 family and reports that it exceeds V4 Pro on performance, speed, and task completion time (DeepSeek positions it as the cost-efficient tier of the V4.1 family and reports that it exceeds V4 Pro on performance, speed, and task completion time.).\n\nResults for the base (pretrained) model are mixed. V4.1-Flash-Base leads on MMLU-Pro, HumanEval and GSM8K. It trails both predecessors on AGIEval, BBH and MATH, and V4-Pro keeps a wide lead on SimpleQA-Verified. Cross-size comparisons are also complicated: backbone parameters are 284B for V4-Flash, 1.6T for V4-Pro and 552B for V4.1-Flash, so the comparison is not like for like on size."
    },
    {
      "heading": "Availability and licensing",
      "text": "The weights are published openly. The model is released with MIT-licensed open weights. Model checkpoints are available on Hugging Face. NVIDIA's model card describes the model as ready for commercial or non-commercial use. It is also offered through DeepSeek's own API and through third-party hosts. Prices differ by provider; one listing on LM Studio, for example, shows input at $0.30 and output at $1.20 per million tokens (Input$0.30 Cached$0.03 Output$1.20). DeepSeek stated that API prices have been reduced accordingly with the release."
    },
    {
      "heading": "Limitations and open questions",
      "text": "Several points remain unsettled. The headline benchmark numbers are self-reported, and the sources reviewed contain relatively little independent evaluation. Performance on some tasks is weaker than in earlier models, as the base-model comparison above shows. The very large total parameter count also means that self-hosting requires substantial multi-GPU infrastructure, which is why most users access the model through hosted APIs. Finally, the terminology used across providers (V4.1 Flash, DeepSeek-V4.1-Flash, deepseek-flash) is not uniform, which can complicate comparisons between sources."
    },
    {
      "heading": "See also",
      "text": "- KV cache\n- Transformer architecture\n- Multimodal learning\n- Context window"
    }
  ],
  "links": [
    "large-language-model",
    "deepseek",
    "transformer-architecture",
    "multimodal-learning",
    "context-window"
  ],
  "sources": [
    {
      "title": "Change Log",
      "url": "https://api-docs.deepseek.com/updates/"
    },
    {
      "title": "DeepSeek V4.1 Flash",
      "url": "https://lmstudio.ai/models/deepseek-v4.1-flash"
    },
    {
      "title": "deepseek-v4.1-flash Model by Deepseek-ai",
      "url": "https://build.nvidia.com/deepseek-ai/deepseek-v4.1-flash/modelcard"
    },
    {
      "title": "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression",
      "url": "https://arxiv.org/pdf/2609.19969"
    },
    {
      "title": "[2609.19969] DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression",
      "url": "https://arxiv.org/abs/2609.19969"
    },
    {
      "title": "DeepSeek-V4.1-Flash: Model Overview & Integration",
      "url": "https://deepinfra.com/blog/deepseek-v4-1-flash-model-overview-integration"
    },
    {
      "title": "DeepSeek V4 Flash",
      "url": "https://lmstudio.ai/models/deepseek-v4-flash"
    },
    {
      "title": "DeepSeek V4.1 Flash - API Pricing & Benchmarks",
      "url": "https://openrouter.ai/deepseek/deepseek-v4.1-flash"
    },
    {
      "title": "DeepSeek V4.1 Flash Fast",
      "url": "https://www.baseten.co/library/deepseek-v41-flash-fast/"
    }
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}