🚨 Tesla China Reportedly Integrates DeepSeek Into In-Car Voice Assistant SystemTesla has integrated Chinese large language models from ByteDance and DeepSeek into its in-car voice assistant for customers in China, according to company documents.For voice command functions such as navigation, media playback and temperature control, Tesla relies on ByteDance’s Doubao model, while interactive AI conversations are handled by the startup DeepSeek.The integration is confined to China, where strict data localizati ...

人工智能迎来两则重磅利好。 首先是，政策层面的利好。8月21日，安徽省人民政府印发《打造通用人工智能产业创新和应用高地若干政策（2.0版）》 ，其中提出了9条具体政策举措，加速大 模型等人工智能技术赋能千行百业，推动安徽省率先进入通用人工智能时代。 在产业层面，DeepSeek今日宣布，DeepSeek-V3.1正式发布。 受此消息影响，DeepSeek概念股——每日互动尾盘直线拉升，截至收盘，大涨13.62%。DeepSeek在文 章中提到，DeepSeek-V3.1使用了UE8M0 FP8 Scale的参数精度。DeepSeek官微在置顶留言里表示，UE8M0 FP8是针对即将发布的下一代国产芯片设计。 安徽重磅发布 8月21日，安徽省人民政府印发《打造通用人工智能产业创新和应用高地若干政策（2.0版）》（以下简称《若干政策》），其中提到，充分发挥财政资金等引导作 用，撬动保险、信贷、基金等社会资本投向通用人工智能产业。加快运营总规模不低于200亿元的省人工智能产业主题基金，以参股方式支持市国资平台、经营主 体等设立通用人工智能领域子基金，满足企业和项目资金需求。建立健全保护改革、鼓励探索、宽容失误、纠正 ...

Core Insights - DeepSeek officially launched DeepSeek-V3.1 on August 21, featuring significant upgrades in its architecture and performance [1]. Group 1: Major Changes in DeepSeek-V3.1 - The new hybrid reasoning architecture allows the model to support both thinking and non-thinking modes simultaneously [2]. - Enhanced thinking efficiency enables DeepSeek-V3.1-Think to provide answers in a shorter time compared to DeepSeek-R1-0528 [2]. - Improved agent capabilities through post-training optimization have led to better performance in tool usage and agent tasks [2]. Group 2: API and User Experience Enhancements - The official app and web model have been upgraded to DeepSeek-V3.1, allowing users to switch freely between thinking and non-thinking modes via a "deep thinking" button [2]. - The DeepSeek API has also been upgraded, with deepseek-chat corresponding to non-thinking mode and deepseek-reasoner to thinking mode, expanding context to 128K [2]. - The API Beta interface now supports strict mode function calling to ensure outputs meet schema definitions [2]. Group 3: Performance Metrics - DeepSeek-V3.1 has shown significant improvements in multiple search evaluation metrics, outperforming R1-0528 in complex search tests requiring multi-step reasoning and expert-level multidisciplinary challenges [2]. Group 4: Technical Adjustments - DeepSeek-V3.1 utilizes UE8M0FP8Scale parameter precision and has made substantial adjustments to the tokenizer and chat template, resulting in noticeable differences from DeepSeek-V3 [3]. Group 5: Market Reaction - Following the announcement of DeepSeek-V3.1, DeepSeek concept stock Daily Interaction (300766) experienced a sharp rise in the late trading session [3].

Core Insights - The article highlights the significant advancements of DeepSeek V3.1, particularly in its ability to handle long contexts and improve programming capabilities, which positions it as a leading open-source model in the industry [1][3][4]. Performance Breakthroughs - DeepSeek V3.1 has achieved a breakthrough in context processing, expanding its context window to 128K tokens, doubling the previous version's capacity, allowing it to handle approximately 100,000 to 130,000 Chinese characters [1]. - The model's enhancements in memory management and attention mechanism have resolved issues related to context loss and fragmented responses in long text processing [1]. Application Scenarios - The model's 128K context capability significantly improves efficiency in legal document review and academic paper summaries, allowing for the input of complete lengthy documents while maintaining logical coherence and detail accuracy [2]. - In developer scenarios, the model supports large codebase dependency analysis and technical document parsing, demonstrating superior context retention and solving previous issues of output loops and information fragmentation [2]. Programming Capabilities - DeepSeek V3.1 has made comprehensive advancements in programming, redefining the performance boundaries of open-source programming models [3]. - In benchmark tests, it scored 71.6% in the Aider Polyglot multi-language programming assessment, outperforming competitors and showing improved accuracy in Python and Bash code generation [4]. Cost Efficiency - The model has achieved a significant cost reduction, with the average cost for completing typical programming tasks being only $1.01, which is 1/68 of closed-source models [7]. - This cost advantage is expected to disrupt the development processes of small and medium enterprises, promoting a shift towards localized, high-efficiency, and low-barrier programming tools [7]. Enhanced Agent Capabilities - DeepSeek V3.1 has improved its tool usage and function calling capabilities, transitioning from "cognitive" to "execution" roles, enhancing its task processing abilities [8]. - The model's compatibility with existing APIs reduces migration costs and enhances cross-platform collaboration efficiency [9]. Reliability and Development Efficiency - The introduction of the Beta version of Strict Mode ensures high accuracy in output formats, particularly in sensitive fields like finance and healthcare, achieving a 99% accuracy rate in data structure compliance [10]. - The model's template-based tool calling reduces integration time by 50%, significantly improving development efficiency [11]. Vertical Capabilities and Practical Applications - The model demonstrates high efficiency in code generation and repair tasks, with costs significantly lower than closed-source competitors [14]. - In enterprise DevOps processes, it automates the generation of deployment scripts, achieving a cost reduction of 1/30 compared to using other models [15]. API Pricing Adjustments - Starting September 6, 2025, DeepSeek V3.1 will adjust its API pricing strategy, with input prices set at 0.5 yuan per million tokens for cache hits and 4 yuan for misses, while output prices will be 12 yuan per million tokens [16]. - Despite some increases in single-call costs, the overall cost-effectiveness remains competitive due to improved token efficiency and faster inference speeds [17].

Core Insights - DeepSeek has announced the release of DeepSeek-V3.1, which utilizes UE8M0 FP8 Scale parameter precision [1] - The V3.1 version has made significant adjustments to the tokenizer and chat template, showing clear differences from DeepSeek-V3 [2] - DeepSeek's official WeChat account indicates that UE8M0 FP8 is designed for the upcoming next-generation domestic chip [3]

Core Insights - DeepSeek officially announced the release of version 3.1 on August 21, featuring significant upgrades including a hybrid reasoning architecture and improved response efficiency [1] - The pricing for API calls has increased, with new rates effective from September 6, including a removal of night-time discounts [2] - The market reacted positively to the news, with shares of Daily Interactive (300766.SZ) rising by 13.62% to 47.98 CNY per share [4] Group 1: Product Updates - DeepSeek V3.1 introduces a hybrid reasoning architecture that allows for flexible switching between thinking and non-thinking modes, enhancing agent capabilities [1] - The model utilizes UE8M0 FP8 Scale parameter precision and has undergone significant adjustments to the tokenizer and chat template, marking a clear distinction from DeepSeek V3 [1] - The foundational model of V3.1 has been retrained with an additional 840 billion tokens, and both the foundational and post-training models are available on Huggingface and MoDa [4] Group 2: Pricing Changes - The API call pricing has been adjusted, with input prices set at 0.5 CNY per million tokens for cache hits and 4 CNY per million tokens for cache misses, up from 2 CNY for V3 [2] - The output price has increased to 12 CNY per million tokens, compared to 8 CNY for V3 [2] Group 3: Market Reaction - Following the announcement, shares of Daily Interactive surged, attributed to its perceived connection with DeepSeek, as it was rumored to hold a 14.50% stake in the DeepSeek development team [4] - Daily Interactive clarified that it does not hold any equity in DeepSeek or its associated companies, despite previous speculation [4]

Group 1 - DeepSeek officially announced the release of version 3.1 on August 21, featuring significant upgrades including a hybrid reasoning architecture and improved response efficiency [1] - The new version utilizes UE8M0FP8Scale parameter precision and has made substantial adjustments to the tokenizer and chat template, showing clear differences from version 3 [1] - DeepSeek has adjusted its pricing for API interface calls, with input prices increasing to 0.5 yuan per million tokens for cache hits and 4 yuan for cache misses, while output prices rose to 12 yuan per million tokens [2] Group 2 - The foundational model of DeepSeek V3.1 underwent extensive retraining, adding a total of 840 billion tokens, and both the foundational and post-training models are available on Huggingface and Modao [4] - Following the announcement, shares of Daily Interactive (300766) surged, closing at 47.98 yuan per share with a daily increase of 13.62% [4] - Daily Interactive, established in 2010, provides data intelligence products and services, and there were rumors of its ownership stake in DeepSeek through its subsidiary, although the company later clarified it does not hold any equity in DeepSeek or its associated companies [7]

Core Insights - The release of DeepSeek V3.1 is positioned as a significant step towards the "Agent Era," featuring a hybrid reasoning architecture that allows the model to switch between fast responses and longer reasoning processes [1] - The new model reduces token generation by 20% to 50% compared to its predecessor, enhancing response speed and lowering usage costs [1] - V3.1 improves throughput efficiency and energy performance, laying the groundwork for large-scale applications [1] - The model demonstrates enhanced capabilities in programming tasks, showing improved execution and stability in real-world environments [1] - In complex search tasks, V3.1 exhibits advanced retrieval and integration abilities, outperforming previous models in multi-disciplinary challenges [1] Business and Ecosystem Strategy - DeepSeek adopts a "dual-track" strategy, continuing to offer API services while adjusting pricing and eliminating night discounts starting September 6 [2] - The base model and post-training versions of V3.1 have been open-sourced on Hugging Face and MoDa [2] Technical Specifications - V3.1 utilizes UE8M0 FP8 Scale parameter precision, aligning with the upcoming generation of domestic chips, which may require specific software adaptations for optimal performance [4] - The release appears to be a direct competitor to GPT-5, with both models supporting long contexts and complex task handling, while offering flexible base model calls and cost structures [4]

DeepSeek unveiled an update to an older model that it says surpasses the seminal R1 on key benchmarks, keeping the Chinese startup in the game while the industry awaits its next flagship offering https://t.co/sOrQY9ug5n ...

2025.08. 21 本文字数：1056，阅读时长大约2分钟 作者 | 第一财 经 刘晓洁 8月21日，业界千呼万唤的R2模型没来，但DeepSeek官方正式发布了新模型V3.1。从命名来看这 或许不是一次大的版本更新，更像是前一代DeepSeek-V3模型的小版本迭代。 在X上，DeepSeek将V3.1称为"我们迈向智能体时代的第一步"（our first step toward the agent era）。本次升级主要有三大亮点，其中包括更强的 Agent能力、混合思考模式和更高的思考 效率。 官方表示，通过后训练优化，新模型在工具使用与智能体任务中的表现有较大提升。在编程智能体、 搜索智能体测评中， V3.1 相比之前的 DeepSeek 系列模型都有明显提高。 | Benchmarks | DeepSeek-V3.1 | DeepSeek- | DeepSeek- | | --- | --- | --- | --- | | | | V3-0324 | R1-0528 | | SWE-bench | 66.0 | 45.4 | 44.6 | | Verified | | | | | SWE-ben ...