The Era of “Deep Training” for LLMs: The Paradigm Shift from Fast Thinking to Slow Reasoning

1. Industry Context: What Does 99 Days of “Re-Training” Really Mean?

For much of the large language model era, attention centered on parameter scale and the accumulation of pretraining data. Over the past year, however, a significant technical shift has emerged: reinforcement learning in the post-training stage is becoming the core variable that determines a model’s intelligence ceiling. A cloud vendor’s latest flagship model reportedly underwent 99 days of RL-based re-training, a duration almost equivalent to training a model from scratch. This is not a routine iteration. It signals a fundamental change in the training paradigm. Models no longer rely solely on memorizing static datasets. Instead, they reconstruct their cognitive capabilities dynamically through interaction with an environment and reward feedback.

From a technical evolution standpoint, the industry has entered the era of deep training. Traditional pretraining gives a model language ability. Deep reinforcement learning gives it the ability to think. The model’s strong global ranking is largely attributed to the introduction of more complex reasoning trajectory rewards and process supervision during the RL stage, rather than merely checking whether the final answer is correct. This training approach enables the model to perform internal reasoning before responding, similar to the deliberate thinking of a human expert.

2. Key Highlights: A Three-Way Breakthrough in Reasoning, Visual Aesthetics, and Coding Benchmarks

2.1 From Fast Thinking to Slow Reasoning: The System 1 to System 2 Transition

The most notable aspect of this model is not any single benchmark score, but its emerging slow reasoning capability. Traditional models usually generate the next token directly based on probability, operating in a System 1 fast-thinking mode. After RL re-training, the model can spontaneously generate an internal chain of thought when facing complex math problems, multi-step logical reasoning, or code debugging. It can even perform self-correction and backtracking.

The key technical factor behind this is the introduction of a process reward model. Unlike conventional reward models that only judge whether the final answer is correct, a process reward model scores each step of the reasoning process. This effectively equips the model with a thinking coach. In testing, when presented with highly difficult competition-level programming problems, the model no longer rushes to output code. It first breaks down the requirements, analyzes boundary conditions, and designs an algorithm framework. Even when errors occur midway, it can revise its approach by simulating execution results. This capability is a qualitative change brought by deep training.

2.2 An Unexpected Lead in Visual Aesthetic Scoring: A Win for Multimodal Alignment

In the intersection of vision and language, the model demonstrates a distinctive sense of composition, color coordination, and layout design. It outperformed a competing product released in the same period on a visual aesthetics benchmark. This is not a matter of mystique. It stems from improvements in the multimodal training strategy. During the RL stage, the training team introduced human aesthetic preferences as a reward signal, combined with joint fine-tuning of a diffusion model and a vision encoder. As a result, the model not only sees images but also understands the aesthetic logic within them.

For developers, this means AI can participate more deeply in UI/UX design, ad creative generation, and short-video storyboard creation. For example, when generating a poster, the model can proactively consider whitespace ratio, contrast between typography and background, and even adjust the emotional tone according to the target audience. These considerations were often overlooked in earlier multimodal models.

2.3 The Engineering Mindset Behind Coding Benchmark Gains

Improved programming capability does not come solely from training on a larger code corpus. It comes from reinforcement learning driven by execution feedback. During RL training, the model was connected to a sandbox compilation environment. Whenever the model generated a piece of code, the system immediately executed it and returned feedback signals such as unit test pass rates and runtime error messages. This combat-drill style of training teaches the model to write code that aligns better with engineering standards: clear variable naming, single-responsibility functions, and complete exception handling.

In specific benchmarks, the model not only solved algorithm problems but also performed strongly in real development scenarios such as code completion and repository-level code modification. This means it has moved beyond solving isolated problems and has begun to understand dependencies across project structures. It can make logical changes across multiple files, which is critical for enterprise-grade applications.

3. API Price Cut by One-Third: The Inference Cost Inflection Point

If improved intelligence is the visible achievement, the significant API price reduction reflects a deeper structural change. The model’s API price was lowered by one-third. This is not simply a promotional discount. It is the inevitable result of technical dividends being released.

Three main drivers are behind the cost reduction:

• Inference engine optimization: Techniques such as paged attention and continuous batching have raised GPU utilization from below 50% to above 80%.
• Lightweight model architecture: Through sparse activation in a mixture-of-experts architecture, each inference pass triggers only about one-tenth of the total parameters, substantially reducing compute consumption.
• Quantization compression breakthroughs: Weights are compressed from FP16 to INT4/INT8 mixed precision, reducing memory usage by 60% while retaining more than 95% of model quality.

For enterprise users, this means the marginal cost of complex tasks such as code review and long-document analysis has dropped sharply. In the past, calling advanced reasoning models for deep analysis was expensive enough to deter many small and medium-sized businesses. Now, lower costs make it feasible to have AI perform multiple rounds of simulated reasoning before providing decision recommendations. This accelerates the shift of AI from assisting human retrieval to independently undertaking analytical tasks.

4. Developer Perspective: How to Evaluate and Use Next-Generation Models Effectively

When integrating models with this level of capability, developers should consider the following strategies.

4.1 Move Beyond Prompt Engineering Toward Task Orchestration

Since the model already possesses deep reasoning ability, developers no longer need complex few-shot examples to coax the model into producing correct answers. A more effective approach is to provide clear goal definitions and constraints. For example, in a code generation task, simply describing “implement an LRU cache that supports high concurrency” allows the model to complete data structure selection and lock strategy design on its own.

4.2 Use the Reasoning Budget Parameter to Control Costs

Although API prices have dropped, deep reasoning inevitably consumes more tokens. Developers should dynamically adjust the reasoning budget according to task difficulty, such as by setting a maximum token limit or an upper bound on reasoning steps. For simple tasks like summarization, the chain of thought can be disabled to obtain sub-second responses. For complex tasks like system architecture design, enabling the full reasoning chain is worthwhile even if it takes several seconds.

4.3 Guard Against Amplified Hallucinations in the Reasoning Chain

Reinforcement learning improves accuracy, but it does not guarantee absolute reliability. When a model reaches an incorrect assumption midway through its reasoning, subsequent steps may confidently proceed further in the wrong direction. For high-value decisions in production environments, it is advisable to add a secondary verification mechanism, such as code execution tests or cross-checking against independent data, rather than blindly trusting the model’s output.

5. Future Outlook: Reinforcement Learning Will Define the Ultimate Height of LLMs

This release sends a clear signal: the basic capability dividend from pretraining is approaching its peak, and post-training reinforcement learning will become the main battlefield for differentiating models. Future competition will focus on three dimensions:

• Reward model granularity: Can automated evaluation systems be built that align more closely with human values?
• Exploration strategy efficiency: How quickly can the optimal solution be found in a vast action space?
• Unified multimodal reasoning: Enabling language, vision, and code to flow seamlessly within a single chain of thought.

For developers and enterprise decision-makers, now is the time to reassess model selection criteria. Parameter size is no longer the sole indicator. The training quality of a model, including RL duration, feedback quality, and alignment level, will directly affect the upper limit of business outcomes. It is advisable to prioritize models that have received targeted reinforcement training in vertical domains such as programming, mathematics, and design, and to establish evaluation sets based on real business scenarios. Final decisions should be grounded in measured results rather than benchmark headlines alone.