智东西（公众号：zhidxcom）
作者 | 程茜
编辑 | 心缘
智东西10月22日报道，今日，科大讯飞发布数个大模型及AI应用新成果，包括端云一体大模型讯飞星火X2.6、语音基座模型SparkAudio 1.0、星火具身多模态交互技术、WinWork数字员工平台、超级AI教师助手讯飞星光等。
其端云一体讯飞星火X2.6基座模型基于全国产算力完成能力升级，包含云端的293-A30B模型及端侧的X2.6-4B模型，后一模型已开源。
云端模型293-A30B支持百万token上下文，输入价格为0.9元/百万tokens，输入缓存命中价格为0.08元/百万tokens，输出价格为3.6元/百万tokens。输入缓存未命中、输出价格比DeepSeek-V4.1-Flash-1017更低。
根据科大讯飞公布的测试结果，星火X2.6-4B模型在信息搜索问答、办公文档处理、日程任务管理、本地表格操作、跨工具协同、多轮工具调用、复杂知识搜索上均超越了其上一代Spark-X2.5-4B以及Qwen3.5-4B、Gemma4-E4B、Qwen3.5-9B模型。
此外科大讯飞还官宣了数个AI硬件，如讯飞AI录音环、讯飞AI眼镜、讯飞AI PC系列等产品。其中灵机N9000 AIPC笔记本最高具备160 TOPS本地AI算力；星火办公一体机提供2240 TOPS AI算力，支持大模型私有化部署。
目前，科大讯飞的产品矩阵已经以讯飞星火基座大模型为底座，在教育、医疗、教育、法律、汽车、机器人等领域实现应用。
科大讯飞创始人、董事长刘庆峰称，今天的AI正成为人类历史上有史以来最具颠覆性的技术革命和产业革命，没有之一。
他认为，全栈自主可控大模型是AI安全性和极致性价比的必然选择，这就需要全栈自主可控的端云一体大模型来解决。未来会有60%的任务在本地跑，30%的任务通过云端协同处理，10%的任务在云端独立处理，这也是科大讯飞此次推出端云一体大模型X2.6的关键原因。
一、基于全国产算力训练，首发端云一体大模型
在科大讯飞公开的基准测试结果中，代码能力方面，星火X2.6在SWE Pro、SWE Multilingual两项测试集中拿下第一，SWE Verified测试集中仅次于DeepSeek-V4.1-Flash 1017；智能体赛道上，其在企业运维类Agent和通用工具调用MCP-Atlas中，都处于第一梯队；在推理能力方面，星火X2.6在AA-LCR v1.1测试集中与MiniMax M3持平。
在训练侧，该模型突破了国产芯片长程任务训练瓶颈，以国产超大规模智算平台飞星二号万卡训练2000B 256K长度数据时间为例，训练时间由4.5个月缩短至1.2个月，超长上下文训练效率提升至原来的4倍。刘庆峰称，星火X2.6就是在过去2-3个月训练出来的。
推理层面，科大讯飞基于国产异构算力的分布式推理技术、首个全国产硬件的四层KV Cache架构，攻克了长上下文场景下容量不足、推理时延高、缓存命中率低等难题，飞星二号国产芯片推理效率达到优化前的4.5倍，KV Cache命中率由63%提升至95.7%。
在端侧模型方面，科大讯飞打造了星火X2.6-4B模型。刘庆峰称，这是因为4B大小是非常适合在机器人、汽车平台、AI PC上部署使用的。其上一代星火X2.5-4B端侧模型从9月1日开源至今，下载量达到256万次，位列全球端侧模型前四。
为提升端侧模型能力，科大讯飞采用了端云一体训练技术。星火采用端侧模型能力感知的自适应迭代后训练算法，根据端侧能力动态选择教师模型、调整训练比重。
基于端侧模型精准路由的“云+端”协同架构，先由端侧模型判断任务交给哪种方式完成：纯端办、纯云办、还是由端云协办。这之中的一个关键问题就是端云路由是否准确，科大讯飞引入基于评语模型的端云路由协同训练机制，将端云路由准确率由81.2%提升至96.1%。
在端云一体的星火X2.6模型之上，科大讯飞推出了两款产品：星火Loomy面向日常办公，帮助用户写报告、做PPT、分析数据，让日常工作轻松完成；星火AStudio面向专业办公，帮助用户写代码、搞研究，让复杂工作自主推进。
现场演示人员扮演一个创业公司的总经理，让星火AStudio可以在不联网的情况下，帮用户查看最近各大区的周报、邮件，然后分析经营情况并进行整理。模型在检索内容后，会生成可视化表格、三项待办事项、风险等，还会提供相应的对应邮件链接。
在和云端模型协作的任务中，模型会先对本地数据进行梳理，有需要云端协作任务时，再发起协作申请。在这之前，本地模型会对数据进行脱敏处理。云端处理完后，本地模型会将之前的脱敏信息在此处理，生成完整的内容。
二、语音基座大模型能分析复杂音频，讯飞AI录音环一点就能随时录音、整理
在让AI从数字世界进入物理世界方面，科大讯飞发布了星火语音基座大模型SparkAudio 1.0以及新款AI硬件讯飞AI录音环等。
星火语音基座大模型SparkAudio 1.0，基于全国产算力训练。现场SparkAudio 1.0分析出了一段复杂音频中的背景音乐、男性声音、机械运转声、自行车铃声等诸多声音。
这一模型背后有三项技术创新，分别为半监督音频编码器带来20%的效果提升，动态分段端到端建模使响应时间降低15%，双范式协同解码使成本降低60%。
该模型已经在讯飞输入法16.0、讯飞听见、讯飞智能办公本、讯飞翻译机等产品中上线。其中，讯飞智能办公本升级后，支持手写标注说话人姓名，后续发言自动对应；抢话、多人交叠发言时也能区分说话人；识别中英混合、专业术语。
科大讯飞还发布了一款新的AI硬件讯飞AI录音环，重量为35g，贴合手机磁吸设计。用户只需点击录音环上的录音按钮，就是随时随地记录。录音结束后，其会通过WiFi 6高速传输，2小时录音约10秒传至手机、2分钟转成文字，还能智能区分说话人、生成AI会议纪要、提炼重点及待办事项，并支持多语种同声转译。
该设备目前已在电商平台开售，售价1299元。
在多模态交互方面，科大讯飞主要提升了座舱、眼镜等智能终端理解需求的能力。
讯飞AI眼镜具备显示能力，重量为40g，支持90dB展会噪声抑制，8米远距拾音。这一设备能翻译120+语种，可以连接PC与云端智能体，支持远程指挥电脑端办公工作。该设备售价为4299元起。
讯飞星火智慧座舱SparkAuto Space 3.0的多人交互响应率、多么题啊交互正确率、车外远场交互距离、复杂智能体任务成功率等都大幅提升。
机器人的具身多模态交互会面临开放空间机器人移动降噪识别、多人交互声源及视觉时空对齐、多模态交互及动作联动协同等难题。
基于此，科大讯飞发布了机器人超脑平台V3.0，能实现复杂空间的动态降噪、多目标人的交互意图识别、语音与动作统一的规划与执行、视觉触发的主动交互。
在“7×24小时”AI无人超市场景中，机器人可以结合语音、视觉与位置信息，识别不同顾客的需求，在多人插话或者用户需求反复变化时，也能完成取货、递送与扫码结账等任务。
机器人超脑平台V3.0将向开发者开放，包括星火超脑板、多模态交互模组、具身智能开发套件及具身多模交互Harness。
在为组织提效方面，AI要完成从个人提效到组织提效的跨越，核心在于读懂业务、嵌入业务流程，实现人、数字员工协同作业。
基于此，讯飞推出讯飞星帆WinWork平台。该平台可以通过搭建企业业务本体的方式，沉淀企业的业务规程、业务数据与专家知识库，让数字员工更理解业务；WinWork还能嵌入企业业务流程，支撑数字员工参与完整工作闭环，从自主识别任务、拆解规划、智能分发，到联动各类业务系统执行。
此外，科大讯飞还推出了基于星火端云大模型、国产操作系统的讯飞AI PC系列产品。其中灵机N9000 AIPC笔记本最高具备160 TOPS本地AI算力；星火办公一体机提供2240 TOPS AI算力，支持大模型私有化部署。
三、给老师、学生、律师、医生、患者都配上AI助手
在教育领域，讯飞发布了一款面向老师的超级AI助手讯飞星光。
讯飞星光将备课、教研、家校沟通和事务助理整合到同一工作台，支持教案、动画、情景视频创编、多角色模拟圆桌研讨、规范填报、统计分析与可视化呈现，对内打通智学网，校信等内部通道，对外打通企业微信、QQ、飞书、钉钉等多平台远控通道。
▲帮老师填写家校沟通表
在备课场景中，为了让学生理解电流方向与磁极方向的关系，老师向讯飞星光提出需求，系统生成可旋转观察的三维交互资源，还可结合摄像头手势操作，让学生模拟右手螺旋定则。
▲可旋转观察的三维交互资源
针对学生，科大讯飞AI学习机上线了“探究学”功能。例如蚂蚁爬行最短路径问题，可以通过三维立体观察，展开长方体表面比较不同路线，让孩子先思考、再动手、最后验证。
在医疗领域，讯飞影像云是全球性医学影像AI服务平台，覆盖了全国超过2.6亿人口的区域。医学影像云平台支持影像共享、智能辅诊、质控与报告生成；讯飞晓医是居民身边的AI健康助手，已累计完成AI咨询超过2.2亿次，下载量突破4400万。
在法律领域，科大讯飞推出一站式AI超级法律助手讯飞晓法，覆盖案情梳理、证据审查、诉讼策略、庭审策略和文书起草等任务。
结语：AI落地需精细化调度模型，端云协同是大势所趋
从科大讯飞本次的新模型与系列新品集中发布可以看出，其正基于讯飞星火基座模型，逐步为企业和个人，打造一套直接可交付的商业化AI方案。
其这次首发的端云一体大模型X2.6，核心是对AI任务需求做精细化拆分调度，以此平衡性能、成本、数据安全三者的矛盾，给开发者与企业提供新的落地选择。尤其对教育、医疗、法律这类对数据出境非常敏感的行业来说，这套架构可以契合其私有化、本地部署的需求。
此外，如今大模型行业的竞争，正持续向算力适配、场景落地、数据安全、软硬一体交付等方向延伸，未来只有模型能力、缺少行业闭环与硬件载体的纯底座玩家或被淘汰。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
I'm sorry.
Edition.
On October 22th, Shigeki reported that today, Kodak Z66Zpublish has several LLM and AI applications for new results, including LLMiFlytek fire X2.6, voice base model SpakAudio 1.0, Starfire multi-modular interaction technology, WinWork Digital Workers Platform, Super AI Teacher Assistant iFlytek starlight, etc.
Its main cloud, the iFlytek Fire X2.6 base model, is based on a national calculator completion capability upgrade, consisting of the 293-A30B model of the cloud and the X2.6-4B model of the end, which is open.
The cloud-based model 293-A30B supports millions of tokencontext, with an input price of $0.9/million tokens and cache hit at $0.08/million tokens and an output price of $3.6/million tokens. Enter cache unhit and output prices are lower than DeepSeek-V4.1-Flash-1017.
According to tests published by Big Z66Z, the Starfire X2.6-4B model went beyond its previous generation of Spark-X2.5-4B and Qwen3.5-4B, Gemma4-E4B and Qwen3.5-9B models in information search questions and answers, office document processing, scheduling task management, local table operations, cross-tool synergy, multiple rounds of tool call, complex knowledge search.
In addition, several AI hardware products such as iFlytek AI tapering, iFlytek AI glasses, and iFlytek AI PC series have been announced. Among them, the pilot N9000 AIPC notebook has up to 160 TOPS local AI calculus; the Starfire Office One provides 2240 TOPS AI algorithms to support the privatization deployment of LLM.
At present, the product matrix of Big Z66Z has been applied in the fields of education, health, education, law, automobiles and robots, using iFlytek fire base LLM as the base.
According to Liu Qingfeng, the founder and chairman of the iFlytek, Koda, today ' s AI is becoming one of the most subversive technological and industrial revolutions in human history.
In his view, the control of LLM is an essential option for AI ' s safety and extreme value for money, and this needs to be addressed by a controlled cloud of LLM. In the future, 60 per cent of the tasks will run locally, 30 per cent through cloud synergetic processes and 10 per cent independently from cloud lines, which is the key reason for this roll-out of Big iFlytek cloud Z43ZX2.6.
LLM
In the publicly available benchmark test results of Cod iFlytek, fire X2.6 was first in the SWE Pro, SWE Multilingual tests, second only to DeepSeek-V4.1-Flash 1017; on the Z42Z track, it was in the first tier of the Enterprise Transport Division Agent and General tool callMCP-Atlas; and in the reasoning capability, fire X2.6 was in the AA-LCR v1.1 test level with Mini Max M3.
On the training side, the model breaks through the long-range mission training bottleneck of the national chip, using, for example, the time spent on 2000B 256K data from the super-large-scale national-based smart-counting platform, Flying Star 2, which reduced the training time from 4.5 months to 1.2 months, and increased the efficiency of the super-long context training four times. Liu Qingfeng claims that Starfire X2.6 was trained over the past two to three months.
At the level of reasoning, Big iFlytek, based on distributed reasoning techniques based on the domestic isomer algorithm, the KV Cache four-storey structure of the first national production hardware, overcome the challenges of inadequate capacity under the long context scene, the high latency and the low cache hit rates, and the efficiency of the Flying Star II national production chip was 4.5 times higher, with KV Cache ' s fatality rate rising from 63 per cent to 95.7 per cent.
In the end-side model, Big iFlytek developed the Starfire X2.6-4B model. Liu Qingfeng says this is because the size of 4B is very suitable for deployment on robots, motor platforms, and AI PC. The previous generation of the X2.5-4B side model has been downloaded 2.56 million times from the open source since 1 September, ranking in the top four of the global side model.
In order to upgrade end-side modelling capabilities, the Big iFlytek of the Section introduced cloud-based training techniques. Fires use end-side model capabilities, a self-adapted, iterative training algorithm that selects teacher models and adjusts training weights according to end-side capacity dynamics.
A cloud-by-end synergetic structure based on the precision of the end-side model, with the end-side model first determining which tasks to perform: pure, pure or co-ordinated. A key issue in this context is the accuracy of cloud trails, which were increased from 81.2 per cent to 96.1 per cent, with the introduction of a collaborative cloud route training mechanism based on comment models.
On the basis of the single cloud X2.6 model, Big Z66Z has launched two products: Loomy of the Starfire aims at day-to-day operations, helping users to write reports, do PPTs, analyse data and make their day-to-day work easier; and Astudio of the Starfire looks at professional offices, helping users to write codes, conduct research and allow complex work to advance on its own.
The presenter acts as the general manager of a start-up company, allowing Astudio, the Starflake, to access the most recent weekly newspapers, e-mails in the regions without a connection, and to analyse and organize the business. After searching for content, the model generates visualized tables, three to-dos, risks, etc., and corresponding mail links.
In the task of collaborating with the cloud model, the model first combes local data and then initiates a request for collaboration when the cloud task is required. Prior to this, local models will desensitize data. When the cloud is finished, the local model will process the previous dissensitization information here to generate the full content.
Two, voice base LLM can analyze complex audio, and iFlytek AI audio sets can be recorded and processed at any moment.
In terms of bringing AI from the digital world to the physical world, Big Z66publish took out the Starfire Voice Base LLMSparkAudio 1.0 and the new AI hardware iFlytek AI audio sets.
Starfire voice base LLMSparkAudio 1.0, based on national computing training. The scene of SparkAudio 1.0 analyzes a complex set of background music, male voices, mechanical sounds, bicycle bells, etc.
There are three technological innovations behind this model, which result in a 20 per cent increase in the performance of semi-supervised audio encoders, a 15 per cent reduction in response time due to dynamic segment-to-end modelling and a 60 per cent reduction in cost due to double-model synergetic decodering.
The model is already online in products such as iFlytek input method 16.0, iFlytek, iFlytek smart office book, iFlytek translator etc. Among them, with the upgrading of iFlytek smart office books, the name of the speaker is supported by hand-written labels and subsequent statements are automatically matched; the speaker can also be distinguished from the speaker when he or she has multiple speeches; and the Chinese and English combinations are identified.
Big Z66Z also publish has a new AI hardware iFlytek AI soundring, weighing 35g, which fits the magnetometer design. Users simply need to click the audio button on the tape, which is always recorded. After completion of the recording, it is transmitted via WiFi 6 high-speed, with two hours of recording being sent to a mobile phone for about 10 seconds, two minutes of text, smartly distinguishing speakers, generating AI minutes, extracting focus and to-dos, and supporting multilingual simultaneous translation.
The equipment is currently being sold on the electrician platform for $1299.
In terms of multimodular interactions, Big iFlytek has mainly enhanced the ability of smart terminals such as cabins, glasses and so forth to understand demand.
iFlytek AI glasses are capable of display, weighing 40 g, supporting the 90dB exhibition noise inhibition and 8 m distance to pick sound. This equipment translates 120+ languages, connects PC to cloud-side AI agent and supports remote command of computer-based operations. The equipment was sold at a price of $4299.
The multi-person interactive response rate of SparkAuto Space 3.0, the issue-like interactive correctness rate, the distance away from the vehicle, the complex AI agent mission success rate, etc. have increased dramatically.
The multimodular interaction of robots is faced with the challenges of open space robots moving noise recognition, multiple interactive sound sources and visual space-time alignment, multimodular interaction and action-linking.
On this basis, Big Z66publish has built a robotic hyperbrain platform, V3.0, capable of dynamic noise reduction in complex space, multi-targeted interactive intent recognition, integrated voice and action planning and execution, and visual activation.
In the “7 x 24 hours” AI no-man's-market scenario, robots can combine voice, visual and location information, identify the needs of different customers, and perform such tasks as pick-up, delivery and billing when multiple interventions or user demands change.
The robotic hyperbrain platform V3.0 will be open to developers, including the superb brain panel of the fire, the multimodular interactive module, the smart development suite and the multimode interactive Harness.
In terms of effectiveness for the organization, the core of AI ' s work to complete the leap from individual to organizational effectiveness lies in understanding business, embedding business processes and working with people and digital staff.
On this basis, iFlytek launched the iFlytek star sail WinWork platform. The platform allows for a better understanding of business by digital employees by building business practices, business data and an expert knowledge base; and WinWork embeds business processes to support the participation of digital staff in the full working closed circle, from the self-identification of tasks, dismantling planning, smart distribution, to the implementation of various business systems.
In addition, Big Z66Z has launched a series of products based on the Star-Fire Cloud LLM and the National Production Operations System iFlytek AI PC. Among them, the pilot N9000 AIPC notebook has up to 160 TOPS local AI calculus; the Starfire Office One provides 2240 TOPS AI algorithms to support the privatization deployment of LLM.
iii. I want teachers, students, lawyers, doctors, patients.
In the field of education, iFlytekpublish has a super-AI assistant iFlytek starlight for teachers.
The iFlytek starlight integrates school preparation, teaching, home-school communication and service assistants into the same workstation, supporting teaching, animation, scenario video creation, multi-role simulation round-table workshops, standard filing, statistical analysis and visualization, internal access to information networks, school letters, etc., and external access to multiple platforms such as WeChat, QQQ, flying books, nails, etc.
I'll help the teacher fill out the home page.
In preparation for the course, in order to give students an understanding of the current direction and the magnetic polar direction, the teacher asked for iFlytek light, the system generated three-dimensional interactive resources that could be rotated, combined with camera gestures, and allowed students to simulate right-hand spiral rules.
3D interactive resource for rotational observation
For students, the Big iFlytek AI learning machine has put on-line the "exploration" function. For example, the issue of the shortest path of an ant's crawling can be seen in three-dimensional stereoid observations, where the rectangular surface can be compared with different routes, allowing the child to think first, then move on, and finally verify.
In the area of health care, the iFlytek image cloud is a global medical image AI service platform covering more than 260 million people in the country. The Medical Image Cloud Platform supports video sharing, intelligence support, quality control and reporting; iFlytek is an AI health assistant to the population, who has completed more than 220 million AI consultations and downloaded over 44 million.
In the area of law, iFlytek introduced all-in-one AI superlegal assistant iFlytek, covering tasks such as combusture, examination of evidence, litigation strategy, trial strategy and drafting of documents.
Conclusion: AI landed in a finely refined schedule model, with cloud synergetic power Trends
As can be seen from the concentration of publish in the new model and series of new products of this major iFlytek, it is building a direct deliverable commercial AI program for businesses and individuals based on the iFlytek Fire Base Model.
The core of its initial cloud, LLMX2.6, is to fine-tune the split movement of AI mission requirements in order to balance performance, cost, data security and provide developers with new landing options. In particular, for industries such as education, health and law, which are very sensitive to data exits, the architecture can match the requirements of privatization and local deployment.
In addition, competition in the LLM industry today is continuing in the direction of calculus adaptation, scene location, data security, hard and hard delivery, and in the future only model capacity, lack of a pure bottom player in the industry closed loop and hardware carriers, or being eliminated.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.