腾讯混元

Platform introduction:

腾讯混元The big model is the core AI base built by Tencent to solve the pain points of "fragmentation of AI capabilities, insufficient adaptation of Chinese scenarios, and difficulty in implementing enterprise-level applications" and positioned itself as a "general intelligent infrastructure." Relying on the Hybrid Expert Model (MoE) structure and 256K ultra-long context capabilities, it has outstanding performance in fields such as Chinese understanding, mathematical logic, and code generation. It can not only provide individual users with services such as content creation, knowledge question and answer, but also provide customized services for enterprises. AI solutions-such as supporting Tencent Meeting AI assistants to improve meeting efficiency, helping Tencent advertising mine potential needs of users, and realizing full-link empowerment "from basic models to scenario applications". It is the core technical pillar of Tencent's AI ecosystem.

Core functions:

1. Five core AI capabilities

   • Multiple rounds of dialogue and understanding : Supports long-text memory and context connection, can accurately understand complex intentions, and adapt to customer service consultation, professional question and answer scenarios (for example, Tencent customer service uses it to achieve multiple rounds of dialogue and reduce business "call rate");
   • Content creation and optimization : Covering literary creation, copywriting polishing, document generation (including tables, formulas, charts), and summary extraction. For example, companies can use it for marketing copywriting and report optimization, and individuals can use it for article creation;
   • Knowledge enhancement and reasoning : Integrate real-time and professional knowledge to solve factual and time-sensitive problems, and at the same time have mathematical and logical reasoning capabilities (such as mathematical calculation and scientific analysis). The Mixed TurboS version further improves the accuracy of reasoning through long and short thinking chains;
   • Code generation and development : Supports code generation, debugging and optimization in multiple programming languages (Python, Java, etc.), helps developers improve development efficiency and adapts to enterprise technology research and development scenarios;
   • Multimodal generation : Achieve "text → image" and "text →3D" generation, support style customization (such as realism, cartoon), and provide visual material support for design, marketing and other fields.

2. Layered model version system

   • Mixed Origin TurboS: The base flagship fast thinking model combines long and short thinking chains, focusing on improving the effects of mathematics, scientific calculation and code generation, and adapting to scenarios with high requirements for response speed and reasoning accuracy;
   • Deep thinking model : Focus on complex task disassembly and in-depth analysis, suitable for scenarios that require fine logical derivation (such as industry report writing, complex problem diagnosis);
   • Hybrid reasoning model : Balancing "speed" and "precision", adapting to most common scenarios (such as daily dialogue, basic content creation), and taking into account efficiency and cost.

3. Full scene application support

   • Document Scenario : Provide full-process assistance of "creation-polish-review", support the generation of tables, formulas and charts, and adapt to the needs of corporate reports, academic documents, personal copywriting, etc.;
   • Meeting Scenarios : Empowering the Tencent Meeting AI Assistant to realize meeting questions and answers, real-time summary, and to-do sorting, and improve meeting efficiency (such as automatically extracting meeting core resolutions and reducing manual recording costs);
   • Marketing and advertising scenarios : Assist Tencent in creating advertising materials, building smart shopping guides, tapping users 'potential business interests, and improving marketing transformation efficiency (such as generating personalized advertising copy to match user preferences);
   • Enterprise service scenarios : Support Tencent customer service, smart shopping guides and other services, solve user problems through multiple rounds of dialogue, and reduce enterprise labor operating costs.

4. Ecosystem and tool linkage

   • Exclusive application carrier : Tencent Yuanbao (an agent creation and distribution platform based on the Mixed Origin Model) provides diversified intelligent services (gradually opening up) to provide individual users with a lightweight AI experience;
   • Enterprise-level access capabilities : Supports API/SDK access to existing systems (such as CRM, customer service platforms, and content management systems) of enterprises to achieve rapid implementation of AI capabilities without the need for enterprises to build their own large models;
   • Security and compliance guarantee : Build a full-process security protection system around the life cycle of large models. It has passed Internet information service algorithm filing and large-scale pre-training model testing certificates to ensure data and application security.

Usage scenarios:

• Improvement in corporate office efficiency : An Internet company uses the Mixed Origin Model to assist in writing quarterly business reports, automatically generates data charts and core conclusions, and shortens the report writing time from 3 days to 1 day;
• Customer service optimization : An e-commerce platform is connected to the mixed power, and the customer service AI can answer user order inquiries and after-sales questions through multiple rounds of dialogues. The "call rate" of manual customer service is reduced by 30%;
• Marketing content creation : A FMCG brand uses mixed energy to generate Short Video scripts and social media copywriting. At the same time, through it to tap potential user needs, the advertising click-through rate increased by 25%;
• Personal learning and creation : Students use mixed elements to help organize course notes and answer mathematical and mathematical problems; self-media creators use it to polish articles and generate graphic and text materials, which improves content output efficiency by 50%;
• Conference management upgrade : Enterprises use Tencent Meeting AI assistant (based on mixed essence) to realize real-time summary of meetings, automatically organize to-do items and assign responsible persons, and the efficiency of follow-up meetings is improved by 40%.

Applicable population:

• Enterprise users : Enterprises in the Internet, e-commerce, finance, education and other industries need to use AI to improve office efficiency, optimize customer service services, and enhance marketing effects (such as Tencent product cooperation companies and third-party corporate customers);
• Developers and technical teams : R & D teams that need to access large model capabilities can integrate mixed capabilities into their own products (such as APP, enterprise systems) through API/SDK to quickly realize the implementation of AI functions;
• Content creators : Self-media people, copywriters, designers, etc., need AI to assist in content creation and material generation (such as graphics and text, code) to improve creative efficiency;
• Ordinary individual users : Individuals (such as students, professionals) who have study, office, and daily Q & A needs to obtain lightweight AI services through Tencent Yuanbao and other carriers;
• Service practitioners : customer service, meeting organizers, marketing personnel, etc., need AI tools to assist in completing professional work (such as customer service questions and answers, meeting summary, and marketing material production).

Unique advantages:

1. Leading the way in Chinese and scene adaptation : Deeply cultivate the Chinese context and perform outstandingly in Chinese understanding, expression and cultural adaptation, far exceeding some overseas models and more in line with the needs of domestic users and enterprises;
2. Strong technical performance : MoE structure is used to improve performance and reduce reasoning costs. The 256K ultra-long context supports long text input (such as complete documents and long dialogues). The mixed-source TurBoS version leads the precision in the fields of mathematics and code;
3. Deep integration of Tencent's ecosystem : Seamless linkage with products such as Tencent Meeting, Tencent Advertising, and Tencent Customer Service, while empowering external companies to achieve a closed-loop "basic model-scenario application" with high implementation efficiency;
4. Security compliance is guaranteed : The full-process security protection system and official filing qualifications solve the concerns of enterprises and individuals on data security and compliance, and is suitable for large-scale commercial applications;
5. Balance between cost and experience : Provide free call quotas, support prepaid and post-paid models, adapt to the cost needs of users of different sizes (individuals, small, medium and micro enterprises, large enterprises), and take into account cost performance and service quality.