On January 30, iFLYTEK held the Spark Cognitive Model V35. Upgrade press conference. Liu Qingfeng, Chairman of iFLYTEK, and Liu Cong, Dean of the Research Institute, officially released the iFLYTEK Xinghuo V3. based on the first national computing power training5. The seven core competencies have been comprehensively improved, and the ability of mathematics, language comprehension, and voice interaction exceeds GPT-4 Turbo.
Artificial general intelligence can not only be deeply and widely used in China's major fields, but also we will stand on a new level in source technology innovation and the underlying capabilities of large models. Liu Qingfeng said.
Based on the national production computing power platform training iFLYTEK Xinghuo v35. The seven major capabilities have been comprehensively improved.
On October 24 last year, iFLYTEK joined hands with Huawei to announce the official launch of the "Feixing-1". In more than 90 days after it was launched, iFLYTEK Xinghuo started the large-scale model training with larger parameters against GPT-4, bringing the iFLYTEK Xinghuo V3 on January 305. Upgrade release.
The first national open large model iFLYTEK Xinghuo v3 based on the training of national computing power5. It has been comprehensively upgraded in seven aspects: language comprehension, text generation, knowledge question answering, logical reasoning, mathematical ability, ability and multimodal ability. Among them, the language comprehension and mathematics ability exceeds that of GPT-4 Turbo, ** reaching 96% of GPT-4 Turbo, and multimodal understanding reaching 91% of GPT-4V.
The large model newly empowers the human-computer interaction experience in the era of the Internet of Everything, and the super-anthropomorphic synthesis effect is amazing. iFLYTEK Spark v35 Not only in the demonstration of semantic understanding, instruction following, and multi-round dialogue, but also in emotional perception and anthropomorphic synthesis.
I heard that Erbin is particularly hot this year, and as a small potato in the south, I really want to play it. Why don't you introduce what's interesting in Northeast dialect? ”
In the practical demonstration session, Liu Cong, Dean of iFLYTEK Research Institute, and iFLYTEK Xinghuo V35. Live interaction, iFLYTEK Spark v35. Amuse the audience with an authentic Northeast dialect. In addition to humor, iFLYTEK Spark v35 quickly customized a travel strategy for Liu Cong, and urged him to buy tickets quickly, and tickets for the Spring Festival were in short supply.
It can not only help users bring solutions, but also bring emotional interaction as a friend who "knows cold and hot", and the ultra-high anthropomorphism makes the large model more humane.
The Spark voice model was officially released.
The first batch of 37 mainstream languages surpassed OpenAI Whisper V3
Large models bring new opportunities for the development of voice technology. Liu Qingfeng emphasized that making machines have the ability to learn, reason and make decisions is the main work of cognitive models. "To put it simply, with the help of large models, we make a speech have richer attributes, including language, content, prosody, timbre, and emotion. ”
The speech recognition effect of the first batch of 37 mainstream languages such as Chinese, English, French, and Russian of the Xinghuo speech model exceeds that of OpenAI Whisper V3, and in terms of multilingual speech synthesis, the first batch of 40 languages of the Xinghuo voice model has an anthropomorphism of more than 83%.
The Xinghuo voice model not only helps international communication, but also "versatile" more scenarios to empower practical applications. Liu Qingfeng introduced that in scenarios such as automobiles, customer service, families, and companion robots, the Xinghuo voice model has more places to play, bringing human-computer interaction changes. For example, empowering automobiles, the interactive experience of intelligent cockpit, intelligent cockpit, intelligent navigation, and first-class control will be further optimized; Industries such as companion robots, shopping guide robots, auxiliary diagnosis robots, smart homes, and wearable devices will also be further improved with the empowerment of voice models.
Xinghuo Open Source-13B was officially released to deeply adapt to domestic computing power.
Sharing sources and developing better software is the key reason why developers, universities, and enterprises are keen on open source for independent research and development, and it is also an important way to build the first developer ecology. At the meeting, the first open-source model based on the national industrialized computing platform "Feixing No. 1" - Xinghuo Open Source-13B was officially released.
Liu Qingfeng revealed that the Xinghuo open source model has formed a differentiated advantage in technology. Xinghuo Open Source-13B ranks among the best in a number of well-known public evaluation tasks, and in typical enterprise scenarios such as text generation, language understanding, text rewriting, industry Q&A, machine translation, etc., through in-depth research and optimization in the fields of learning assistance and language understanding, the practicability has been greatly improved, and it is more handy in dealing with complex natural language tasks.
Based on the training of "Feixing-1", the Xinghuo open-source large model is optimized for full-stack domestic adaptation, which is simple and easy to use, leading the application effect in scenarios, and the training strategy is optimized for the ultimate computing power of Ascend, with a training efficiency of 90% of that of A100. This is not only a further in-depth optimization of Ascend AI hardware, but also demonstrates the determination and ability of domestic computing power to catch up with the international advanced level.
Only when more scenarios are implemented in open source can ecological cooperation be better enhanced. At present, Huawei's Shengsi open source community has officially launched the open source version of the Xinghuo large model-13b, which is completely free for academic and enterprise research, enhancing academic cooperation and industry exploration.
Upstream news reporter Zhang Hao