Since September of the previous year when OpenAI revolutionized the model training paradigm with the GPT-o1 model, domestic companies have started to meet the industry expectations, initiating a substantial wave of o1-like model development. Here is the core content formatted for a WordPress blog post:
Updates in January 2025
In the first month of 2025, domestic o1-like models have seen a flurry of updates, with releases from notable players including “The Six Little Tigers” – Dark Side of the Moon, Step Star, and the independently operating DeepSeek.
DeepSeek’s Model Release
DeepSeek officially launched the DeepSeek-R1, a model that aligns with the performance of OpenAI’s o1 formal version, while also open-sourcing the model weights. Test results reveal that DeepSeek-R1 matches OpenAI-o1-1217 in various tasks and even outperforms it slightly on some test sets. Key points include:
- DeepSeek also open-sourced DeepSeek-R1-Zero, exploring the technical feasibility of training large language models through reinforcement learning alone.
- In terms of pricing, DeepSeek continued its strategy of low cost, offering prices significantly below that of OpenAI o1.
Dark Side of the Moon’s K1.5
Dark Side of the Moon introduced K1.5, a multimodal thinking model positioned as a “multimodal o1”. Official data shows its capabilities are on par with GPT-4o and Claude 3.5 Sonnet across various domains. Features include:
- K1.5 employs reinforcement learning and multi-stage training to achieve multimodal capabilities.
- In reinforcement learning, K1.5 uses methods like “length penalty” to suppress response length.
Step Star’s Step R-mini
Step Star launched the Step Reasoner mini experimental version, a reasoning model with super-long inference capabilities. Key points are:
- It currently competes with OpenAI o1-preview and o1-mini on test sets.
- Step Star emphasizes its balance of “arts and sciences,” capable of handling a variety of tasks.
Analysis of the Overall Situation
The wave of domestic o1-like models following the trend is a testament to the industry’s dynamism. Technically, however, domestic models have yet to incorporate more complex technologies, and it remains uncertain whether this is the key to enhancing reasoning abilities. Additionally, with OpenAI’s rapid pace of updates, domestic large model companies face significant competitive pressures.
以下 is the formatted content for WordPress:
The Domestic o1-like Model Surge
In 2025, domestic o1-like models have undergone intensive updates, including DeepSeek’s R1, Dark Side of the Moon’s K1.5, and Step Star’s Step R-mini. These models, each with their unique performance attributes, have embraced technologies like reinforcement learning. However, domestic models still face technical challenges and competitive pressures. Below is the detailed content:
The DeepSeek-R1 has demonstrated impressive performance across multiple tasks, coupled with affordable pricing. The K1.5 from Dark Side of the Moon stands out as a multimodal o1, boasting versatile multimodal capabilities. Step Star’s Step R-mini highlights its balance in arts and sciences. Overall, domestic models must accelerate innovation to navigate the competitive landscape.
Here is the summary of the article content:
In a narrative that blends scientific rigor with a touch of humanistic flair, the domestic modeling landscape is painted as one of relentless pursuit and quiet resilience. As models like DeepSeek-R1, K1.5, and Step R-mini emerge, they not only challenge the status quo but also embody the spirit of innovation in the face of adversity.
