DeepSeek V4.1 Flash Hugging Face model card and technical report

DeepSeek V4.1 Flash Hugging Face model card and technical report

• ID: 19819179
• 原文链接: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
• 作者: DeepSeek-AI
• 日期: 2026-09-10
• 平台: manual
• 来源类型: product
• 标签: deepseek, v4.1-flash, huggingface, technical-report, open-weights, field-note
• 质量评分: 4/5
• 抓取时间: 2026-09-11T15:54:30+00:00
• 抓取状态: ok

中文导读

DeepSeek-V4.1-Flash 的 Hugging Face 页面是模型权重模型卡和技术报告的主要入口，与官方新闻中的 API 发布互补DeepSeek 新闻页把 HF 链接列在模型开源部分，并说明会支持社区进行推理适配扩大部署范围；本地调研材料还把它与 API 名 deepseek-flash552B CED MoEArtificial Analysis 分数和价格口径分开记录该条目用于锚定可下载模型与技术报告，而非二手评论

为什么值得关注

官方新闻讲 API 与价格，HF 页面锚定模型卡权重和技术报告；两者合在一起才是 V4.1 Flash 的完整发布面

English summary

The Hugging Face page is the canonical model-card and technical-report entry point for DeepSeek-V4.1-Flash, complementing the API launch announcement. It anchors the downloadable/open-weight side of the release, while the DeepSeek news post covers API routing, pricing and official claims around the 552B CED MoE design.

图片

• https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true
• https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true
• https://img.shields.io/badge/%F0%9F%A4%96%20Chat-DeepSeek%20V4.1-536af5?color=536af5&logoColor=white
• https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white
• https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white

抓取内容（opencli-first）

deepseek-ai/DeepSeek-V4.1

发布时间: 2026-09-10T04:09:09

原文链接: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

Edit model card

#deepseek-v41-flash-pushing-the-limits-of-kv-cache-compressionDeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

• * *

https://www.deepseek.com/https://chat.deepseek.com/

https://huggingface.co/deepseek-aihttps://twitter.com/deepseek_ai

https://huggingface.co/deepseek-ai/LICENSE

**Technical Report** 👁️

#introductionIntroduction

We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively.

Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads. SWA Bounded Replay reconstructs missing SWA KV states by replaying only the most recent _n_\_win tokens, avoiding the need to persist SWA KV to SSD and reducing the persistent KV cache footprint to roughly 1/8 of that of DeepSeek-V4-Flash.

Compressed Sparse Attention 2 (CSA2). DeepSeek-V4.1-Flash uses CSA2, which assigns each attention layer one of three static modes — Full, Reindex, or Reuse — to share main KV and indexer K across layers and reuse Top-K sparse-attention indices. In the decoder, a Hierarchical Sparse Indexer further restricts later indexing layers to a candidate pool constructed by the first Full Mode layer, bounding deeper indexer cost independently of context length. Combined with FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels), these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.

Additional architectural components include Single-Pass mHC (revised residual-stream mixing with an efficient Mega-mHC kernel), Engram conditional memory (196B parameters, sparsely accessed via token-based lookup), and DSpark speculative decoding (semi-autoregressive draft generation with confidence-scheduled verification). The model uses 1 shared expert and 384 routed experts per MoE layer, activating 6 routed experts per token.

Multimodal architecture. A vision encoder (DeepSeek-ViT, trained from scratch with 2D-RoPE and 3×3 pixel-unshuffle downsampling) and a two-layer MLP projector convert images into visual embeddings, processed jointly with text embeddings from the start of language-model pre-training.

Pre-training. DeepSeek-V4.1-Flash is trained from scratch on a multimodal corpus comprising 45T tokens, with sparse attention trained at a sequence length of 64K and context extended to 1M tokens at 34T tokens.

Post-training. The post-training recipe follows the standard SFT → RL → on-policy distillation (OPD) paradigm without algorithmic modifications. All substantive changes lie instead in the data pipeline: large-scale automated synthesis of agent tasks and environments with progressive scaling of data, tasks, and rollouts. The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.

_Figure 1. (a) Performance of DeepSeek-V4.1-Flash and counterparts on agentic benchmarks. (b) Global KV cache size per token (bytes) across generations of DeepSeek models. DeepSeek-V4.1-Flash achieves approximately 4-fold and 437-fold reductions relative to DeepSeek-V4-Flash and DeepSeek-V1, respectively._

#evaluation-resultsEvaluation Results

#base-modelBase Model

All base models are evaluated in our internal framework under the same evaluation settings. Scores within 0.3 of each other are considered equivalent.

| 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 |

#instruct-modelInstruct Model

DeepSeek-V4.1-Flash supports a continuously controllable reasoning effort from 1 to 100. All instruct results below use the maximum effort setting (reasoning_effort=100). Evaluations use temperature=1.0, top_p=0.95.

For code agent benchmarks (Terminal-Bench 2.1/3.0/4.0, DeepSWE v1.1, NL2Repo-Bench, ProgramBench), the model is evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window. To align with official setup requirements, the mini-SWE harness is used for DeepSWE v1.1, and the Claude Code harness for SEC-Bench Pro. Visual agent benchmarks (Chartography, BabyVision, ZeroBench) use the Claude Code harness with a 512k-token context window. Agent's Last Exam and AutomationBench use their official scaffolds. All agentic evaluations use temperature=1.0, top_p=0.95.

#comparison-with-frontier-models-max-reasoning-effortComparison with frontier models (Max reasoning effort)

| Benchmark (Metric) | Opus-5.0 | GPT-5.6 Sol | K3 | GLM-5.3 | DS-V4-Pro | DS-V4-Flash | DS-V4.1-Flash | | :-- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | | Reasoning | | | | | | | | | GPQA Diamond (Pass@1) | 93.4 | 94.1 | 92.9 | 88.1 | 92.4 | 89.9 | 90.9 | | HLE (Pass@1) | 56.3 | 44.5 | 43.5 | 42.0† | 42.7† | 37.8† | 36.8 (39.1†) | | Codeforces (Rating) | — | — | — | — | 3348 | 3289 | 3471 | | MathArena Apex (Pass@1) | — | — | 65.6 | — | 65.3 | 58.6 | 65.6 | | Agentic | | | | | | | | | Terminal-Bench 2.1 (Pass@1) | 89.1 | 88.8 | 88.3 | 88.2 | 87.9 | 82.7 | 90.6 | | Terminal-Bench 3.0 (Pass@1) | 43.3 | 34.4 | 17.7 | 28.3 | 11.8 | 7.6 | 30.0 | | Terminal-Bench 4.0 (Pass@1) | 51.8 | 39.9 | 12.6 | 37.9 | 12.4 | 7.0 | 31.2 | | DeepSWE v1.1 (Resolved) | 74.0 | 73.0 | 67.5 | 66.9 | 62.7 | 54.4 | 74.2 | | ProgramBench (Almost@1) | 37.0 | 23.0 | 17.5 | 19.0 | 15.5 | — | 20.3 | | NL2Repo-Bench (Score) | 75.3 | 56.8 | 58.0 | 58.0 | 61.5 | 54.2 | 64.0 | | CyberGym (Pass@1) | — | 84.5 | 80.0 | 84.5 | 83.3 | 76.7 | 88.1 | | SEC-Bench Pro (Pass@1) | — | 74.3 | — | — | 56.4 | 30.9 | 62.8 | | ExploitGym (Pass@1) | 22.1 | 33.7 | — | 15.0 | 5.4 | 1.8 | 15.3 | | HLE w/ tools (Pass@1) | 63.6 | — | 59.8 | 62.5 | 60.0 | 51.5 | 63.9 | | AutomationBench (Pass@1) | 50.3 | 45.8 | 46.7 | 48.8 | 43.2 | 37.7 | 54.8 | | Agent's Last Exam (Pass@1) | 28.6 | 26.7 | 27.6 | 28.5 | 25.7 | 25.2 | 31.8 | | Chartography w/ tools (Pass@1) | 84.0 | 79.9 | 68.1 | — | — | — | 78.9 | | BabyVision w/ tools (Pass@1) | 94.1 | 88.9 | 85.7 | — | — | — | 89.6 | | ZeroBench-main w/ tools (Pass@5) | 52.0 | 53.0 | 41.0 | — | — | — | 49.0 |

_† Text-only subset of HLE._

#performance-across-agent-scaffolds-deepswe-v11-and-terminal-bench-21-max-reasoning-effortPerformance across agent scaffolds (DeepSWE v1.1 and Terminal-Bench 2.1, Max reasoning effort)

All scaffolds use N=8 samples per task on DeepSWE v1.1 and N=3 on Terminal-Bench 2.1, with Linux containers, temperature=1.0, top_p=0.95, a 1M-token context limit, and max\_steps=500 per agent. Terminal-Bench 2.1 is evaluated 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 |

#prompt-encodingPrompt Encoding

This release does not include a Jinja-format chat template. The `encoding` folder contains a self-contained Python reference implementation (encoding.py) with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved image content.

For production use, we additionally release deepseek-recipe, a set of Rust libraries with Python bindings that provides the same prompt format as a maintained, protocol-aware toolkit. It converts Messages, Chat Completions, and Responses API requests into the Conversation format, encodes them into DeepSeek V4 and V4.1 prompts or token IDs, and parses model output back into complete or streamed responses — covering thinking, tool calls, images, and generation settings. Model inference, tool execution, and HTTP transport are left to the caller.

#minimal-inferenceMinimal Inference

Please refer to the `inference` folder for instructions on weight conversion and running inference locally.

Recommended sampling parameters:

| Parameter | Value | | :-- | :-- | | temperature | 1.0 | | top_p | 0.95 or 1.0 | | context_window | 1M tokens | | max_tokens | ≥ 256K |

#reproducing-deepswe-benchmark-resultsReproducing DeepSWE Benchmark Results

The `evaluation` folder contains step-by-step instructions for reproducing the DeepSWE v1.1 benchmark results, covering both the dsh-minimal agent and the official mini-swe-agent. The patch required to integrate dsh-minimal with Pier is also included there.

#licenseLicense

This repository and the model weights are licensed under the MIT License.

#citationCitation

@misc{deepseekai2026deepseekv41flash,
      title={DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression},
      author={DeepSeek-AI},
      year={2026},
}

#contactContact

If you have any questions, please raise an issue or contact us at service@deepseek.com.

Downloads last month

75,774

Safetensorshttps://huggingface.co/docs/safetensors

Model size

763B params

Tensor type

BF16

F32

F8\_E4M3

I8

Files info

Model tree for deepseek-ai/DeepSeek-V4.1-Flashhttps://huggingface.co/docs/hub/model-cards#specifying-a-base-model

Finetunes

4 models

Quantizations

https://huggingface.co/models?apps=llama.cpp&other=base_model:quantized:deepseek-ai/DeepSeek-V4.1-Flash "Use with llama.cpp"https://huggingface.co/models?apps=lmstudio&other=base_model:quantized:deepseek-ai/DeepSeek-V4.1-Flash "Use with LM Studio"https://huggingface.co/models?apps=jan&other=base_model:quantized:deepseek-ai/DeepSeek-V4.1-Flash "Use with Jan"https://huggingface.co/models?apps=ollama&other=base_model:quantized:deepseek-ai/DeepSeek-V4.1-Flash "Use with Ollama"

32 models

Spaces using deepseek-ai/DeepSeek-V4.1-Flash 6

Collection including deepseek-ai/DeepSeek-V4.1-Flash

[

10 items • Updated 1 day ago • 901

](https://huggingface.co/collections/deepseek-ai/deepseek-v4)

Evaluation resultshttps://huggingface.co/docs/hub/eval-results

• Idavidrein/gpqa · Diamond View evaluation results leaderboard

90.9

• harborframework/terminal-bench-2.1 · Terminalbench 2 1 View evaluation results leaderboard

https://huggingface.co/datasets/harborframework/terminal-bench-2.1?eval_result=deepseek-ai%2FDeepSeek-V4.1-Flash&leaderboard_task_id=terminalbench_2_190.6 \*

• datacurve/deep-swe · Deep Swe View evaluation results leaderboard

https://huggingface.co/datasets/datacurve/deep-swe?eval_result=deepseek-ai%2FDeepSeek-V4.1-Flash&leaderboard_task_id=deep_swe74.2 \*

• cais/hle · Hle
• default View evaluation results

36.8

• With tools; harness not specified in the model card. View evaluation results

https://huggingface.co/datasets/cais/hle?eval_result=deepseek-ai%2FDeepSeek-V4.1-Flash&leaderboard_task_id=hle63.9 \*

• harborframework/terminal-bench leaderboard
• +2 more

Like your first model

Support deepseek-ai and save this model in your likes history.

Dismiss

Inference providers allow you to run inference using different serverless providers.

View evaluation results shared by the community

Ranked #1 overall

DeepSeek Harness, Minimal mode, 1M-token context.

View evaluation results shared by the community

Ranked #1 overall

Reported as 'DeepSWE v1.1'; official mini-swe-agent harness.

View evaluation results shared by the community

Ranked #1 overall

With tools; harness not specified in the model card.

Obsidian evidence excerpt

Master Research · DeepSeek-V4.1-Flash

合并自 00-local-context + 01-dump-official-api-hf + 02-dump-aa-third-party。

一句话

2026-09-10 DeepSeek 发布 V4.1-Flash（新架构族最小档、原生多模态、API 名 deepseek-flash、MIT 开源）；同日 Artificial Analysis 给出 Index 40、AutomationBench-AA 69%、一等价 $0.30/$1.20、约 $0.27/Index 任务。读材料必须把官方合同行和第三方分栏拆开；另外 9/14 V4 Pro 是否下线 以现网 api-docs 注脚为准（已改为继续提供 Pro）。

官方合同行（可写进「DeepSeek 说了什么」）

| 项 | 内容 | 来源 | |----|------|------| | 发布时间 | 2026-09-10 | 官帖 / news | | 定位 | 新架构族最小；原生视觉 | 官帖 | | 体量 | 552B MoE；输入激活 8B / 输出 16B | news | | 架构名 | Causal Encoder–Decoder | news | | KV | 相对上代 1/4 HBM、1/8 SSD | news | | API 名 | deepseek-flash | news + docs | | 旧 Flash | 下线；旧 id 路由到 V4.1-Flash 按 Flash 价 | docs | | 上下文/输出 | 1M / 最大 384K | pricing | | Vision | Flash 有；Pro 无 | pricing | | 峰谷价 | 闲时=高峰 50%；高峰 UTC 工作日 01–04 与 06–10 | pricing | | Flash 高峰价 | in miss $0.30 / cache $0.006 / out $1.20 | pricing | | 开源 | HF MIT；Tech Report PDF | HF + 6/6 帖 | | V4 Pro 路由 | news 曾写 9/14 04:00 UTC 起 pro→Flash 且按 Flash 价；现网 docs (2) 改为 9/14 后继续提供 Pro、计费不变 | news vs docs 以 docs 为准 | | 热度 | 主帖 ~2.67 万 likes / ~537 万 views（晚间） | fxtwitter |

AA 第三方行（只标 AA，不并进官方口径）

| 项 | 内容 | |----|------| | Index | 40（称超 V4 Pro 0813；DeepSeek 侧 AA 榜叙述新旗舰） | | Pro 对照 Index | V4 Pro 0813 (max) 约 36（AA provider 页） | | AutomationBench-AA | 69%（= Astra 叙述；657 tasks / 39 SaaS） | | AA-LCR v1.1 | 84% | | GDPval-AA v2 | 1632 Elo（自 1468） | | Terminal-Bench v4.0 | 27%（不采帖内倍数修辞） | | 价（一等） | $0.30 / $1.20；cache $0.006；闲时再半 | | $/Index task | ~$0.27 | | 啰嗦度 | ~89k tokens / Index task | | 参数叙述 | 帖内 552B 与 763B total 并存；HF total≈763B |

工程读法（主稿判断）

1. 做路由表：至少四行——官方能力声明 / AA Index / 工作流类 AutomationBench / 单价与 $/task（含啰嗦度） 2. 账单：agent 多轮看 cache hit 价 + 峰谷窗；别只比 output 标价 3. 别把 AA「超过 Pro」写成 DeepSeek 官网 slogan 的唯一来源——官网有自家 benchmark 图与「多方测试」句，AA 是独立套件 4. 9/14 前若文案仍写「Pro 将全部变成 Flash」，先打开 pricing 注脚 (2)

明确不写进主结论

• 本机未跑 benchmark / 未测 tok/s
• 未 OCR 官网 benchmark 图逐格分数
• Cursor / Edge0 / 金融 Work 不并题