2025 03 11 The word count of this article is 3,146, and the reading time is about 5 minutes Introduction Rocky Mountain Technology opened up on March 11, recording 3 boards in 5 days.2025 Year-to-date, Rocky Mountain Technology’s share price has been soaring, and as of the close of trading on March 11, compared with the year’s low, the share price is up by nearly 100 . Author: Wang Yuanli
2025 03 11 The word count of this article is 3146, and the reading time is about 5 minutes Introduction Rocky Mountain Technology opened up on March 11, recording 5 days and 3 boards.2025 year-to-date, Rocky Mountain Technology shares soared all the way, as of the close of March 11, compared with the year’s lows, the share price rose by nearly 100 .
2025 03 11
This article has a word count of 3,146 and a reading time of about 5 minutes
Introduction Rocky Mountain Technology opened up on March 11, recording three boards in five days.2025 Year-to-date, Rocky Mountain Technology shares have soared, and as of the close of trading on March 11, compared to the low point of the year, the share price is up nearly 100 .
Author: Wang Yuanli
AI intelligent body robot concept brain-computer interface automatic driving Take a number of popular concepts of rocky mountain science and technology 002195 SZ March 11 opening halt, recorded 5 days 3 board. Since the beginning of 2025, Iwama Science and Technology shares have been soaring, as of the close of trading on March 11, compared with the low point of the year, the share price rose nearly 100 .
Rocky Mountain Technology formerly known as two three four five, early main business for the Internet information services, core products include 2345 browser 2345 decompression software 2345 king of tools such as software.
With the decline of the traditional Internet business, 2-3-45 in 2023, the official use of artificial intelligence business as a strategic catalyst for a new round of growth, the layout of the field of artificial intelligence in the field of intelligent driving Brain intelligence and generative AI, and renamed Iwashan Technology.
Behind the soaring stock price is a sharp decline in performance. Performance snapshot shows that rocky mountain technology 2024 is expected to realize net profit of 28 million yuan 42 million yuan, a year-on-year decline of 91 56 87 34 . The company explained that one of the reasons for the sharp decline in performance is that it has increased its investment in research and development in the artificial intelligence segment, which has increased by about 100 million yuan year on year. According to Rocky Mountain Technology’s 2024 semi-annual report, the artificial intelligence business currently accounts for only 10 of total revenue .
Shanghai Rockcore Digital Intelligence Artificial Intelligence Technology Co. Rockcore Digital Intelligence for short Rockcore Digital Intelligence is a subsidiary of Iwama Science and Technology established in June 2023, which is mainly responsible for the research, development, and application of the field of artificial intelligence generative AI.
According to the company’s official microblogging information on March 5, the end-side big model Yan model developed by Rockcore Numerical Intelligence has made some progress in commercialization. Can the AI business of Iwashan Technology become a new performance growth point for the company in the future?
The cost of training does not exceed 3 million, Yan1 3 big model Comparable to Llama3?
According to reports, Yan big model is independently developed by Rockcore Digital Intelligence, focusing on the research and development and application of natural language understanding and human-computer interaction technology.
Rockcore Numerical Intelligence was founded in June 2023, according to the data of the regular report of Iwama Technology, as of the mid-2024 report, Iwama Technology has invested a total of 16.2 million yuan in Rockcore Numerical Intelligence through its own funds, with a shareholding of 99 .
More than half a year after its establishment, Rockcore Numerical Intelligence released the Yan1 0 model to the public in January 2024, and launched the Yan1 2 model in July 2024, which was upgraded to the Yan1 3 model in September of the same year. It is reported that Yan model is a non-Transformer architecture general-purpose natural language model. According to the test data disclosed by Vivante, Yan model has higher performance than Transformer architecture models with the same parameters.
The Yan model was evaluated using the 5-shot test method with 4 typical natural language models for 5 basic tasks. The data shows that the average score of Yan1 3 3B model is higher than that of Meta’s open source Llama3 8B model, and its scores are higher than those of Llama3 8B model in pronoun disambiguation and contextual reasoning in WinoGrande’s main model, and in common sense reasoning and natural language reasoning in HellaSwag’s main model.
Source: Rocky Mountain Technology 2024 Half-Yearly Report Rockcore Digital Intelligence
Compared with the Transformer architecture, where the training cost of large models can easily reach tens of millions or even hundreds of millions, and the total investment is not more than $20 million, can the Yan model developed by Rockchip Numerical Intelligence really achieve the same performance?
In an interview with China Business News, Liu Fanping, CEO of Rockcore Numerical Intelligence, said that the training cost of the 3B multimodal model of Yan architecture, for example, does not exceed RMB 3 million from zero pre-training to fine-tuning and alignment.
Why can achieve such a low training cost, Liu Fanping said, mainly because of the infrastructure level of improvement and innovation, and the current version does not have autonomous learning, the future of the architecture with autonomous learning training costs should be lower.
According to the company, RockAI’s Yan model is based on a bionic neuron-driven selection algorithm, and the model can be activated in partitions based on the type of learning and scope of knowledge. The synchronous learning concept, on the other hand, allows the model to update its knowledge in real time during training and reasoning, a learning method similar to the working mechanism of the human brain, which is able to quickly adapt to new situations and demands.
As to whether the Yan model outperforms Llama3, it is difficult to replicate the Yan model because it is not an open-source model.
Some experts in the field of large models told CBN that benchmarking is not the only standard for measuring the ability of large models, and more attention needs to be paid to the application of the model’s landing effect.
Yan model application in the field of robotics still has challenges
Compared to the Transformer big model even answer a hello need full parameter activation, Yan model brain-like partition activation mechanism in the deployment of low-computing power equipment has a natural advantage, Yan1 3 multimodal big model can be realized in the drone robot PC and other types of terminals on the deployment of non-destructive. Liu Fanping said.
According to Rocky Mountain Technology’s interactive easy news on February 27, the company’s Rockcore Digital Intelligence cooperated with a person-type robotics company, deployed the self-developed Yan1 3 big model on a robot platform based on Intel Core i3 chips, and delivered a robot equipped with a big model to its third-party customer.
As to whether the Yan model’s performance meets the requirements of humanoid robots for a big model of embodied intelligence, the manufacturer of a humanoid robot body told the reporter that it has only seen the model’s performance on the end-side semantic big model, while the robot body needs a big model of embodied intelligence, that is, a fusion multimodal big model with hearing, seeing, thinking and acting.
The embodied intelligence model needs to have generalized scene perception and command understanding, then perform generalized action understanding and planning, and finally generate fine action training. Currently, Yan model has certain advantages in the first step of generic scene perception and command understanding, but it has not demonstrated its modeling capability in the last two steps.
The person in charge of a head of the humanoid robot enterprise told CBN that most of the head of the humanoid robot enterprises have adopted a self-research approach to layout the large model of embodied intelligence. For the large model of embodied intelligence, pure semantic understanding is not useful, the core difficulty lies in the need for real scenes, and humanoid robot end-side action data combined with training, so the pure software level of architectural innovation can not solve the application of large models in the field of humanoid robotics problems.
From the Yan model of large-scale commercialization of landing attempts, Rockchip Digital Intelligence also chose AIPC, rather than the field of robotics.
On March 5, Iwanshan Technology subsidiary Rockcore Digital Intelligence released information in its official microblogging said that in the 2025 Mobile World Congress, a head of the sea brand released AIPC products equipped with its self-developed Yan architecture big model, opening a new era of end-side AI scale landing, but due to commercial confidentiality, the name of the partner was not disclosed.
Can it bring incremental performance?
For the above and a brand cooperation AIPC products when can bring specific performance income, the reporter asked the relevant person in charge of rocky mountain technology, rocky mountain technology, said that due to confidentiality agreements can not be disclosed, the specific progress of the company concerned about the relevant announcements.
According to rocky mountain technology 2024 performance snapshot data, its 2024 is expected to achieve net profit of 28 million yuan 42 million yuan, down 91 56 87 34 .
Wind data show that the Internet information service business of Rocky Mountain Technology has continued to decline in revenue in recent years, sliding from 1.665 billion yuan in 2019 to 485 million yuan in 2023, and realizing Internet information service revenue of 228 million yuan in the first half of 2024, continuing to decline year-on-year 6 8 .
However, it is noteworthy that its AI business realized revenue of 34.38 million yuan for the first time in the first half of 2024, accounting for 10 16 of the total revenue. Since the launch of its artificial intelligence strategy in 2023, in addition to the field of large models, Rocky Mountain Technology has also laid out the fields of automatic driving and brain-computer interface.
On April 2, 2024, Iwashan Technology, through its wholly-owned subsidiary Shanghai Ruifeng Intelligent Technology Co., Ltd. invested in Newmai Technology, an intelligent driving company, at a total pre-investment valuation of RMB 1.98 billion, for an amount of RMB 675 million, with a shareholding of 26,12 .
In August 2024, Iwasan Technology established Shanghai Iwasi Brain-like Artificial Intelligence Research Institute Co., Ltd. and as of the mid-year report of 2024, it has invested a total of RMB 18 million, accounting for 100 percent of the shares. Shanghai Iwasi Brain-like Artificial Intelligence Research Institute is oriented to the prospective research in the field of brain science and artificial intelligence with the EEG data and EEG model as the core technological base.
Iwashan Technology can layout cutting-edge technology in multiple lines, but also with its low debt ratio and sufficient cash. According to the data of Rocky Mountain Technology’s 2024 third quarterly report, as of September 30, 2024, Rocky Mountain Technology held monetary cash of 1.265 billion yuan, and its gearing ratio was only 2 69 . Its 2024 earnings flash report said that the company continued to increase R&D investment in the AI segment, and R&D expenditures for the AI segment in 2024 increased by about 100 million yuan year-on-year.
A primary market investor in the field of science and technology told the first financial, Rock Core Digital Intelligence Yan model Whether the future can really bring profit to Rock Mountain Technology, need to wait for further commercialization of related products to verify.
WeChat Edit Donkey of the production team Pulling the mill version
Recommended Reading
Leave a Reply