From Base to Reasoning: The Evolution of Open-Source LLM Architecture and Training Methods
This article traces the iterative development of an open-source large language model (LLM) family, analyzing how the team evolved its architecture and training strategies across three major releases. It covers the cost-efficient base model built on a Mixture-of-Experts (MoE) architecture, the subsequent reasoning-focused version driven by a reinforcement learning-first paradigm, and the latest formal release featuring sparse attention mechanisms for long-context scenarios. The discussion distills reusable methodologies for model selection and deployment strategy, offering practical insights for developers and enterprises evaluating open-source LLMs.
1. Why Continuous Iteration Matters
The competitive landscape for open-source large language models has shifted significantly over the past two years. One open-source model team completed three key iteration milestones within roughly a year. First, they released a base model that rivaled top-tier closed-source models at an extremely low training cost. Next came a reasoning-enhanced version that achieved a qualitative leap in complex tasks like mathematics and code generation. After a quiet period of nearly half a year, the team released an intermediate experimental version followed by an official release whose reasoning benchmarks now approach those of leading international closed-source models.
The value of studying this evolution lies in how it illustrates the trade-offs an open-source team makes across four dimensions: cost constraints, architectural consistency, training paradigm innovation, and inference optimization. For developers and enterprise users, this is not just news—it serves as a practical reference for technology selection.
2. Foundational Concepts for Understanding the Evolution
Before analyzing specific changes, it is essential to clarify several concepts that are often confused.
2.1 The Two Stages of LLM Training
Pre-training phase: The model is trained on massive text corpora to produce a base model. The base model functions essentially as a “text continuator”—it predicts subsequent text based on preceding context, but it does not naturally have conversational ability, instruction-following capability, or safety refusal mechanisms.
Post-training phase: Training continues on top of the base model to teach it how to converse with humans. Post-training typically includes two steps:
- SFT (Supervised Fine-Tuning): The model is trained on manually annotated, high-quality question-answer pairs to learn basic conversational formats and instruction following.
- RL (Reinforcement Learning): Reward signals further optimize model behavior, such as using a reward model to evaluate response quality or providing verifiable rewards for correct reasoning.
2.2 Reasoning Models vs. Standard Models
A standard model directly outputs an answer upon receiving a question. A reasoning model, by contrast, first generates an explicit “thinking process” before producing the final answer. The thinking content is often wrapped in special tags, for example:
The final answer is...
This paradigm gives the model a larger “compute budget” before output, significantly improving accuracy on complex tasks. The trade-off is an increased number of output tokens and longer response latency.
3. Phase One: The Base Model—Engineering Efficiency for Higher Performance Ceiling
The core contribution of the first base version is its extreme training efficiency. According to public technical reports, its training cost was on the order of a few million dollars, yet it reached the performance level of contemporary leading closed-source models. This was the result of multiple coordinated optimizations in engineering and algorithms:
- Mixture-of-Experts (MoE) architecture: The model’s total parameter count is very large, but each token activates only a small subset of expert networks, so inference computation is far lower than that of a dense model of comparable scale.
- Multi-token prediction training objective: Each forward pass predicts multiple subsequent tokens, improving sample efficiency.
- Low-precision training: Lower-precision numerical formats reduce memory and communication overhead while maintaining training stability.
The significance of this architecture is that all subsequent versions reuse the same architecture, with changes focused on training methods—a classic “stable architecture, evolving methods” strategy.
4. Phase Two: The Reasoning Version—A Breakthrough in Training Paradigm
The reasoning version released shortly afterward was a training methodology innovation under an identical architecture to the base version. This is the key to understanding the entire evolution path.
4.1 Two Training Variants
- Pure reinforcement learning variant: This approach took the base model directly, skipped SFT entirely, and used pure RL. The model spontaneously developed long-chain reasoning abilities on tasks with verifiable rewards (mathematics, code), but outputs suffered in readability and mixed languages.
- Cold-start variant: This approach first applied a small amount of SFT “cold start” using several thousand high-quality reasoning samples, followed by large-scale RL. It also incorporated SFT data generated by rejection sampling and alignment training focused on readability. The final result outperformed the pure RL variant and more closely matched human usage habits.
4.2 Technical Insights
This phase validated an important conclusion: reasoning abilities can largely emerge from reinforcement learning rather than relying entirely on massive human annotation. This significantly lowered the barrier to training reasoning models and was the direct impetus for the industry’s broader shift toward an “RL-first” paradigm.
5. Phase Three: Sparse Attention and Inference Efficiency Optimization
After a quiet period—during which infrastructure-level adjustments reportedly involved adaptation for domestic AI accelerator chips—the team released both an intermediate experimental version and an official version. The intermediate version did not show standout performance; its real purpose was to pave the way for architectural changes, allowing inference frameworks and deployment tools to complete adaptation in advance.
The core change in the official version was the introduction of DeepSeek Sparse Attention (DSA):
- Problem context: As context length grows, standard full attention incurs quadratic computational costs, making long-context inference increasingly expensive.
- Solution approach: A lightweight indexing mechanism first selects the few tokens from the long context that are most relevant to the current query, then precise attention computation is applied only to these selected tokens.
- Benefits: This significantly reduces computation and memory overhead in long-context scenarios while preserving model quality, thereby lowering inference costs.
- Engineering coordination: Because DSA requires specialized operator support, releasing the experimental version in advance gave community inference engines sufficient time to adapt—a mature practice in open-source ecosystem management.
After the official release, reasoning benchmarks approached the level of leading closed-source models, validating the viability of the “focused architectural optimization plus ecosystem co-adaptation” path.
6. Abstracting the Evolution Path
Reviewing the entire evolution, three reusable methodologies emerge:
- Get the architecture right once; iterate on methods: After the MoE architecture was established in the base version, it was not substantially changed. Subsequent innovations concentrated on training paradigms and attention mechanisms, reducing ecosystem adaptation costs.
- Build capabilities in layers: Base capability comes from pre-training efficiency; reasoning capability comes from the RL paradigm; cost efficiency comes from attention sparsification. Each step solves one core problem at a time.
- Release experimental versions first: Before infrastructure changes, release an experimental version to secure ecosystem adaptation time and avoid compatibility shocks at the official release.
For enterprises and developers, the selection criteria should not be based merely on version numbers, but on the actual capability dimensions these changes represent: whether long-context low-cost inference is needed, whether complex reasoning capabilities are relied upon, and whether the deployment framework already supports the corresponding operators. Understanding the logic of this evolution enables technology decisions that align with your specific use case.
