DeepSeek V4.1 Flash: The Native Multimodal Model That's Breaking Speed Records

查看原文 / View Original 编辑 / Edit

DeepSeek V4.1 Flash: The Native Multimodal Model That’s Breaking Speed Records

DeepSeek V4.1 Flash：打破速度记录的原生多模态模型

How DeepSeek Achieved 420 Tokens/Second Without Sacrificing Accuracy DeepSeek 如何在不牺牲准确性的前提下实现每秒 420 token 的速度

Published: September 9, 2026 | Reading time: 10 minutes 发布日期：2026年9月9日 | 阅读时间：10分钟

The Surprise Beta

突如其来的测试版

On September 8, 2026, DeepSeek quietly launched a beta test for V4.1 Flash, an intermediate model with the identifier deepseek-v4.1-flash-expires-on-0910. The model name itself reveals the urgency—this beta expires on September 10, giving developers just 48 hours to test it. What makes this launch remarkable isn’t just the speed. It’s that V4.1 Flash is DeepSeek’s first native multimodal model, supporting both text and image input/output from the factory, not as an afterthought. 2026年9月8日，DeepSeek 低调开启了 V4.1 Flash 的内测，该中间版本模型的标识符为 deepseek-v4.1-flash-expires-on-0910。模型名称本身就透露出紧迫感——该测试版将于9月10日过期，开发者仅有48小时进行测试。此次发布引人注目的不仅是速度，更在于 V4.1 Flash 是 DeepSeek 首个原生多模态模型，它从底层架构上就支持文本和图像的输入/输出，而非后期补丁式集成。

Performance Benchmarks

性能基准测试

TaskSpeed Improvement
49K Long Context Retrieval5.2x faster
SVG Code Generation6.0x faster
Manacher Palindrome Algorithm4.6x faster
Complex SQL Query Generation5.0x faster
Asyncio Architecture Refactoring3.9x faster
任务速度提升
49K 长上下文检索5.2倍
SVG 代码生成6.0倍
Manacher 回文算法4.6倍
复杂 SQL 查询生成5.0倍
Asyncio 架构重构3.9倍

Peak Performance: 420 tokens/second in long-text reasoning tasks, with end-to-end throughput reaching 409.5 tokens/second. 峰值性能： 在长文本推理任务中达到每秒 420 token，端到端吞吐量达到每秒 409.5 token。

Real-World Test: A user sent a photo of a person in a striped suit. The model correctly identified the striped pattern, avoiding the hallucination issues that plagued earlier vision models. 实测表现： 一位用户发送了一张身穿条纹西装的人的照片。模型准确识别出了条纹图案，避免了困扰早期视觉模型的幻觉问题。

The Architecture: What’s New?

架构解析：有何新意？

Native Multimodal Support 原生多模态支持

Unlike V4 Flash Vision-Exp, which added a visual encoder as an “external plugin” on top of a text-only base, V4.1 Flash integrates text and image processing from the ground up. This means: 与 V4 Flash Vision-Exp 不同（后者是在纯文本基座上添加视觉编码器作为“外部插件”），V4.1 Flash 从底层开始就集成了文本和图像处理。这意味着：

  • Unified Representation: Text and images share the same latent space.
  • 统一表征： 文本和图像共享相同的潜在空间。
  • Better Cross-Modal Reasoning: The model can reason across modalities more effectively.
  • 更好的跨模态推理： 模型能更有效地进行跨模态推理。
  • Lower Latency: No need to coordinate separate encoders and decoders.
  • 更低延迟： 无需协调独立的编码器和解码器。

New Model Structure 全新模型结构

DeepSeek claims V4.1 Flash uses a “new model structure,” but hasn’t released a technical report. Community analysis suggests: DeepSeek 声称 V4.1 Flash 使用了“全新模型结构”，但尚未发布技术报告。社区分析认为：

  • Improved Attention Mechanism: Likely building on CSA/HCA hybrid attention from V4.
  • 改进的注意力机制： 可能基于 V4 的 CSA/HCA 混合注意力机制构建。
  • Optimized Routing: Better expert selection in the MoE architecture.
  • 优化路由： 在 MoE 架构中实现了更好的专家选择。
  • Enhanced Multimodal Fusion: Deeper integration of visual and text features.
  • 增强的多模态融合： 视觉和文本特征的深度集成。

The Engineering Challenge: Why 150 New Engineers?

工程挑战：为何需要 150 名新工程师？

DeepSeek simultaneously announced hiring 150 senior engineers, focusing on: DeepSeek 同时宣布招聘 150 名高级工程师，重点关注：

  • Backend Development: Model research platforms, Agent frameworks, API infrastructure.
  • 后端开发： 模型研究平台、Agent 框架、API 基础设施。
  • Agent Computing: Platform development, low-level optimization, elastic computing.
  • Agent 计算： 平台开发、底层优化、弹性计算。

As Cui Tianyi, DeepSeek Harness负责人, explained: “When quantity increases, complexity explodes exponentially. Data volume, machine/container count, training tasks, evaluation tasks, Agent environments, user count, request volume—all are increasing dramatically. This creates complexity that old backend systems can’t handle.” 正如 DeepSeek Harness 负责人崔天一所解释的那样：“当数量增加时，复杂性会呈指数级爆炸。数据量、机器/容器数量、训练任务、评估任务、Agent 环境、用户数、请求量——都在急剧增加。这带来了旧后端系统无法处理的复杂性。”

The Real-World Impact

实际影响

V4.1 Flash’s speed isn’t just a benchmark achievement. It enables: V4.1 Flash 的速度不仅仅是基准测试的胜利，它还实现了：

  • Real-Time Multimodal Interaction: Users can send images and receive responses in under a second.
  • 实时多模态交互： 用户发送图像后，可以在一秒内收到响应。
  • Complex Agent Workflows: The model can handle multi-step tasks with dynamic environments.
  • 复杂的 Agent 工作流： 模型能够处理动态环境下的多步骤任务。
  • Cost Efficiency: Same pricing as V4 Flash, but with better performance.
  • 成本效益： 与 V4 Flash 价格相同，但性能更优。

Code Example: Multimodal Input

代码示例：多模态输入

from openai import OpenAI
client = OpenAI(
    api_key="your-api-key",
    base_url="https://api.deepseek.com/v1"
)

# Send an image with text prompt
response = client.chat.completions.create(
    model="deepseek-v4.1-flash",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What is the person wearing in this image?"},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/person.jpg"}
                }
            ]
        }
    ],
    max_tokens=500
)

print(response.choices[0].message.content)

The Pricing Advantage

价格优势

Despite the performance improvements, V4.1 Flash maintains the same pricing as V4 Flash. This is significant because: 尽管性能有所提升，V4.1 Flash 仍保持与 V4 Flash 相同的定价。这一点意义重大，因为：

  • Competitive Pressure: Other models are charging more for similar or worse performance.
  • 竞争压力： 其他模型在性能相似或更差的情况下收费更高。
  • Market Disruption: DeepSeek is using price to gain market share.
  • 市场颠覆： DeepSeek 正利用价格优势抢占市场份额。
  • Sustainability: The new architecture is more efficient, allowing lower prices.
  • 可持续性： 新架构效率更高，从而支持更低的价格。

Community reaction: When performance is high and price is low, Liang Wenfeng (DeepSeek’s founder) becomes “梁圣” (Saint Liang). When prices rise, he becomes “梁子” (Little Liang). With V4.1 Flash, he’s back to being 梁圣. 社区反应： 当性能高且价格低时，梁文锋（DeepSeek 创始人）被称为“梁圣”。当价格上涨时，他被称为“梁子”。随着 V4.1 Flash 的发布，他又变回了“梁圣”。

The Beta Limitations

测试版局限性

What We Know: 已知信息：

  • Intermediate Version: This is not the final release; it expires on September 10.
  • 中间版本：这不是最终发布版；它将于 9 月 10 日过期。
  • No Technical Report: DeepSeek hasn’t published detailed architecture documentation.
  • 无技术报告：DeepSeek 尚未发布详细的架构文档。
  • Limited Access: Only available through beta testing, not public API.
  • 访问受限：仅通过内测提供，非公开 API。

What We Don’t Know: 未知信息：

  • Exact model architecture details.
  • 确切的模型架构细节。
  • Full benchmark results across all tasks.
  • 所有任务的完整基准测试结果。
  • Long-term stability and reliability data.
  • 长期稳定性和可靠性数据。
  • Final pricing strategy for the public release.
  • 公开发布的最终定价策略。

Industry Context: The AI Race Is Heating Up

行业背景：AI 竞赛持续升温

In the past month, Chinese AI models have been released at an unprecedented pace: 过去一个月里，中国 AI 模型以史无前例的速度发布：

  • July: Kimi K3 open-sourced, Qwen3.8-Max released.
  • 7月： Kimi K3 开源，Qwen3.8-Max 发布。
  • August: GLM-5.3 released, DeepSeek V4 Pro launched.
  • 8月： GLM-5.3 发布，DeepSeek V4 Pro 上线。
  • September: DeepSeek V4.1 Flash beta, V4.1 Pro expected.
  • 9月： DeepSeek V4.1 Flash 内测，预计 V4.1 Pro 即将到来。

This “weekly release” rhythm is unprecedented in the AI industry. 这种“周更”节奏在 AI 行业中前所未有。

The Competition

竞争格局

ModelRelease DateKey Feature
Kimi K3July 20262.8T parameters, native multimodal
GLM-5.3August 2026Post-training scaling, cyber security
DeepSeek V4 ProAugust 2026Million-token context, DSA attention
DeepSeek V4.1 FlashSeptember 2026Native multimodal, 420 tok/s
模型发布日期核心特性
Kimi K32026年7月2.8T 参数，原生多模态
GLM-5.32026年8月后训练扩展，网络安全
DeepSeek V4 Pro2026年8月百万 token 上下文，DSA 注意力
DeepSeek V4.1 Flash2026年9月原生多模态，420 tok/s

What’s Next?

未来展望

Expected Timeline: 预期时间表：

  • September 10: V4.1 Flash public release (tentative).
  • 9月10日：V4.1 Flash 公开发布（暂定）。
  • September 2026: V4.1 Pro release (expected).
  • 2026年9月：V4.1 Pro 发布（预期）。
  • Late 2026: Next-generation models (speculative).
  • 2026年底：下一代模型（推测）。

What to Watch: 关注重点：

  • V4.1 Flash Public API: When it launches, expect massive adoption.
  • V4.1 Flash 公开 API：发布后预计将迎来大规模采用。
  • V4.1 Pro Performance: How much better than Flash?
  • V4.1 Pro 性能：比 Flash 强多少？
  • Pricing Strategy: Will DeepSeek maintain low prices?
  • 定价策略：DeepSeek 是否会维持低价？
  • Ecosystem Development: Tools, frameworks, and integrations.
  • 生态发展：工具、框架和集成。
← 返回新闻列表 / Back to news list