DeepSeek推出了V4.1-Flash-1017，这是对其Flash模型系列的一次更新1。新版本采用稀疏混合专家架构，基于因果编码器-解码器（CED）设计，在输入时激活8B参数、输出时激活16B参数2。该模型在推理、编码、代理任务和视觉基准上实现了实质性改进1，同时通过压缩KV缓存等优化手段，使缓存内存降低至前一代Flash版本的约四分之一2。
新版本在性能和成本效率方面都有显著进展。任务执行时间和输出长度均得到缩减，单次API调用的综合成本下降3。通过基于自反馈的强化学习训练，模型增强了从零构建完整项目的能力，支持游戏开发、3D建模等场景的项目构建3。该模型具有原生图像理解能力，视觉和文本嵌入从预训练阶段开始联合训练2。
模型权重已在HuggingFace和ModelScope平台以MIT许可证开源3，用户可下载使用。API调用方式与价格保持不变，仍通过deepseek-flash进行访问3。DeepSeek同时为Transformers、vLLM、SGLang等多种推理框架提供了使用说明1。
DeepSeek has released V4.1 Flash-1017, an incremental update to its V4.1 Flash model that delivers substantial improvements across reasoning, coding, agent tasks, and visual benchmarks.12 The new version achieves these enhancements while reducing task execution time, output length, and overall user costs.1
The V4.1 Flash-1017 employs a sparse mixture-of-experts architecture based on a novel causal encoder-decoder (CED) framework.2 During inference, the model activates 8 billion parameters for input processing and 16 billion parameters for output generation, with a backbone network of 552 billion parameters.2 The model incorporates native image understanding capabilities, with visual and text embeddings jointly trained from the pretraining phase.2 A key technical optimization compresses the KV cache to approximately one-quarter of the memory footprint required by the previous Flash generation.2
The model demonstrates performance gains that surpass V4 Pro across speed and task completion metrics.2 Through reinforcement learning training based on self-feedback, the model has been enhanced to build complete projects from scratch, supporting use cases in game development and 3D modeling.3 Benchmark results include scores of 42.6 on HLE (without tools), 77.0 on DeepSWE v1.1, 92.5 on BabyVision (with tools), and 42.3 on Terminal-Bench 4.0.3
The model weights are released under the MIT license and are available for download on HuggingFace and ModelScope.13 API call methods and pricing remain unchanged, with the model continuing to be accessed via the deepseek-flash endpoint.3 Developers using platforms such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy are advised to upgrade to version 0.2.1-rc.1 or later for optimized compatibility.3
评论
还没有评论，欢迎留下第一条。