Why not use deepseek directly for the ideal large vehicle model?
The upgrade of Ideal Classmate AI brings a new experience of deep thinking in vehicle scenarios.
Core content:
1. Ideal Auto OTA 7.2 version update, vehicle-mounted autonomous decision-making reasoning model online
2. Ideal Classmate deep thinking function: voice interaction, instant understanding of needs, personalized solutions
3. The unique advantages of Ideal's self-developed reasoning model and its adaptability to vehicle scenarios
On April 2, Ideal Auto OTA 7.2 brought a major update that made AI "open up" - the in-vehicle autonomous decision-making reasoning model was officially launched. Users only need to turn on the "deep thinking" switch in the vehicle control center, and Ideal students will have the superpower of "double online quotient", which can analyze and answer complex questions, and decide whether to think deeply based on the type of question.
"Help me plan a 3-day, 2-night in-depth travel guide for Beijing with a cost of 2,000 yuan per person. Avoid popular check-in spots and focus on experiencing the old Beijing culture." - When such meticulous needs are thrown to Ideal Classmate, it can not only understand your thoughts in seconds, but also respond within hundreds of milliseconds and quickly provide personalized plans including hutong homestays, niche museums and private restaurants.
AI technology media @Quantum位 also immediately evaluated such a "smarter" Ideal Classmate: "Ideal Classmate has evolved again. The newly added deep thinking switch can combine your verbal prompts to prioritize the actions and reply statements that need to be executed. It is more suitable for vehicle-side scenarios. Simple operations can be executed directly, and complex questions can be answered in detail after thinking. It can be seen from the answers that Ideal Classmate really thought about it, rather than simply searching and aggregating. "
Source video account @QuantumBit
“A good smart cockpit must not only be fast, but also be able to think.” Only through personal experience can you know that such precise and efficient interaction is not something that ordinary in-car voice assistants can do.
At present, the automotive industry is experiencing a "DeepSeek on-board wave". According to statistics from the 21st Century Business Herald, about 20 automakers have announced deep integration with DeepSeek. However, is this cooperation model really suitable for in-vehicle scenarios? Ideal Auto gave a different answer - tailor-made inference models for in-vehicle scenarios. What are the considerations behind this? What are the unique advantages of Ideal's self-developed model? Let's find out through the following questions and answers!
Q1
Why do ideal students think deeply?
Ideal students are connected to the inference model developed by Ideal Auto that is tailored for in-vehicle scenarios . After turning on the "deep thinking" switch, this model can intelligently identify the type of question and independently decide whether to conduct deep thinking based on the type of question.
Q2
Is Ideal's inference model directly connected to DeepSeek R1?
No. Our model is an inference model developed by Ideal Auto specifically for vehicle-mounted scenarios, rather than directly using DeepSeek R1. Although we referenced general inference model data including DeepSeek R1 during the training process, the final product is a dedicated solution that is fully optimized for the characteristics of vehicle-mounted scenarios.
Q3
Why did Ideal choose to develop its own inference model instead of directly using DeepSeek?
First of all, the in-vehicle scenario is different from accessing DeepSeek on mobile phones and PCs. We will have many vehicle control commands and questions that need answers immediately. This type of need can quickly respond to user needs and is not suitable for the reasoning model's way of thinking first and then answering. In this case, the general reasoning model needs to be implemented by manually turning off the "deep thinking" switch. This is a safety hazard when driving, and the experience is very poor.
In addition, in-vehicle scenarios are mainly based on voice interaction, which is more sensitive to response speed (imagine you are talking to someone and he waits dozens of seconds before replying to you each time). The general reasoning model is slow to answer, which is not friendly to the in-vehicle user experience. At the same time, the information density of voice will be lower, and the reply of DeepSeek R1 is too long for us.
The most suitable one is the best, so we created a reasoning model that responds faster, outputs text that is more suitable for broadcasting, and has a "deep thinking" switch that can be always on.
Q4
What are the characteristics of an ideal classmate who is capable of deep thinking?
There are three characteristics in total: First, ideal students who can think deeply have greatly improved their ability to deal with complex problems, and are adept at logical reasoning, investment analysis, strategy formulation, story creation, etc.
In addition, the overall response speed is very fast. Our actual test results on OTA 7.2 show that the delay of the first token of Deepin is only 300 milliseconds.
Finally, you can decide whether deep thinking is needed or not, without having to manually switch the "deep thinking" switch. Ideal students will decide whether deep thinking is needed based on different types of instructions. For example, vehicle control instructions will be executed directly; for complex issues such as strategy planning, deep thinking will be conducted to give a more comprehensive and in-depth response. This ensures that Ideal students' original experience in vehicle control and car travel remains as good as ever , and the answers to complex questions are also richer and more accurate.
Q5
Why is Ideal Classmates’ response so fast?
The speed advantage comes from two aspects: first, the model can independently decide whether deep thinking is needed to avoid unnecessary computing overhead; second, it maintains an extremely fast response speed during deep thinking. We chose the MindGPT model with tens of billions of parameters. Through phased supervised fine-tuning, enhanced reasoning optimization, and preference alignment, we achieved the full performance of DeepSeek R1 while also achieving a faster response speed than R1.
Q6
Can you use specific data to illustrate the speed and effect advantages of ideal students?
In actual tests, the first token delay of Ideal Deep Thinking is only 300 milliseconds. Comparative tests show that our exclusive model performs roughly on par with DeepSeek R1 in vehicle-related datasets, and is also close to R1 in general dataset tests. This is due to the reasoning acceleration strategy we adopted, including technical optimizations such as PD separation, high-performance attention mechanism, and engineering parallel reasoning.
Q7
How can we make independent decisions on whether to think deeply?
The core of this capability is that our model has learned a set of intelligent decision-making mechanisms. We are the first car company to launch a self-developed large-scale in-vehicle model, and we have millions of users, so we have accumulated a lot of experience in in-vehicle interaction scenarios. Based on this experience, we have established a complete distribution of in-vehicle scenario questions and answers, and based on this, we have built an adaptive "fast thinking (short), slow thinking (long)" reasoning training data set. Based on this data set, we continue to train the reasoning model, and through supervised fine-tuning and preference alignment, the model has the ability to make autonomous decisions "fast thinking and slow thinking".
Q8
What challenges did you encounter during model training?
There are two main challenges: one is how to accurately determine the type of problem so that the model has the ability to independently decide whether to think long or short; the other is how to deal with common fuzzy problems in car scenarios to prevent the model from falling into an endless loop of thinking. For the latter, we mined a large amount of fuzzy expression instruction data, built a concise and efficient training data set, and significantly improved the model's processing efficiency on such problems through SFT and reinforcement learning.
Q9
What technological innovations does Ideal have in model training?
We use a multi-stage supervised fine-tuning combined with reinforcement learning training methods, and use advanced algorithmic technologies such as curriculum learning and generative reward models to achieve phased enhancement and efficient iteration of model capabilities. Last year alone, we updated more than 30 versions and continued to optimize the user experience. From the launch of Ideal Classmates in 2019 to the launch of the self-developed large model MindGPT in 2023, we have accumulated rich experience in the field of in-vehicle AI and built a large amount of high-quality training data, which is our core advantage.
In the future, we will adhere to independent research and development to ensure efficient model iteration and bring users a smarter and smoother in-vehicle interactive experience!
