Large Model Capabilities, 2022–2026: Evolution, Evaluation, and Lessons Learned

    A retrospective of how large model capabilities evolved between 2022 and 2026, and what that evolution means for developers and enterprises. It follows the shift from conversational AI becoming a mass-market product in late 2022 to coding agents that deliver complete tasks, traces four capability tracks (open weights, native multimodality, video generation, and software engineering agents), and examines scientific AI's move from predicting protein structures to predicting biomolecular complexes. It then looks at how China's model release wave narrowed into a competition over engineering and deployment, and ends with three questions the industry still needs to answer: what models can reliably do, how generated content should be governed, and what makes large model economics sustainable. It also offers practical guidance on evaluation (task completion rate, hallucination rate, long-context stability) and includes a Python sketch for evaluating coding agents.

    1. Where the Cycle Began: Conversational AI Goes Mainstream

    At the end of 2022, a conversational AI product officially launched and replaced traditional command-driven human-machine interaction with natural language. Within two months, it passed 100 million monthly active users and was regarded as one of the fastest-growing consumer applications ever. It therefore became the widely accepted starting point of this large model cycle. Its significance lay not in leading on any single metric, but in moving “large model capabilities” out of labs and papers and into a product ordinary people could use every day.

    For developers, the key lesson of this stage was that shifts in interaction paradigms often matter more than parameter scale in determining how fast a product spreads. Packaging complex capabilities into a seemingly simple chat box is high-value engineering work in its own right. It involves a full stack of infrastructure, including context management, safety alignment, rate limiting, and cost control.

    2. Four Main Lines of Capability Evolution

    Looking back over four years, capability advances did not come from linearly piling on parameters. Instead, they unfolded along several clear tracks.

    2.1 From Closed Source to Open Weights: A Watershed for the Ecosystem

    In early 2023, weights for a mid-sized model from an international company were accidentally released. The model had been intended for research use by application only, and the leak directly triggered a wave of community fine-tuning and distillation. Comparable vendors followed by releasing model weights under commercial licenses. This laid the foundation for the open-weight ecosystem in Western markets.

    The real value of open weights is that they lowered the barrier to innovation. Developers could run continued pretraining on their own data, instruction fine-tuning, and quantized deployment. They could turn a general-purpose base model into a specialized model tailored to a vertical use case. For small and mid-sized businesses, this also meant much better control over inference costs and data compliance.

    2.2 Native Multimodality: The Default Architecture

    Most early multimodal approaches bolted a vision encoder onto a language model and combined images and text through feature concatenation. Native multimodal models, by contrast, jointly model text, images, audio, and video as unified modalities from the pretraining stage onward.

    This shift has a clear engineering consequence: cross-modal alignment no longer depends on post-processing but forms intrinsically during training. Flagship models are typically split into large, medium, and small tiers, such as ultra-large-scale, general-purpose, and on-device/edge tiers, so teams can balance quality against cost across different deployment environments.

    2.3 Video Generation: From “Frame Sequences” to “World Simulation”

    In early 2024, a video generation technology entered the public eye. Its selling points were minute-long clips and strong spatiotemporal consistency, and it was framed in the narrative as a “world simulator.” It lowered the barrier to creation from specialized software to text descriptions. It also triggered debate over copyright, deepfakes, and content authenticity. For engineering teams, the lesson was clear: before generative content can scale, provenance marking and content moderation pipelines must be built in alongside it.

    2.4 Agents: From “Completion” to “Delivery”

    In 2024, a coding agent marketed as an “AI software engineer” appeared. It could work autonomously in an isolated sandbox: reading documentation, writing code, running tests, and fixing errors. It pushed coding assistance from “completing the next line” to “delivering a complete task.”

    The most important methodological legacy of this stage was a shift in evaluation criteria. A coding agent is no longer judged on whether the code it generates looks right. It is judged on its end-to-end task completion rate on real repository tasks. The controversy around it, especially the gap between demo performance and real-world capability, in turn pushed the industry toward more rigorous evaluation standards.

    A schematic example of the evaluation logic:

    1. # Coding agent evaluation: use task completion rate as the core metric, not generation quality
    2. def evaluate_agent(agent, task_suite):
    3. passed = 0
    4. for task in task_suite: # each task: repo snapshot, problem statement, hidden tests
    5. result = agent.run(task) # read docs -> edit code -> run tests -> iterate on failures
    6. if all_tests_pass(result, task.hidden_tests):
    7. passed += 1
    8. return passed / len(task_suite) # end-to-end task completion rate

    The key design points are:

    • Tasks must come from real engineering work, not artificial examples.
    • Test cases should be hidden from the model to prevent overfitting.
    • Verdicts should be reproducible and regression-friendly.

    3. Scientific AI: From Predicting Structures to Predicting Complexes

    In the same year, a biomolecular structure prediction model made a breakthrough. It expanded its prediction targets from single protein chains to complexes made up of proteins, DNA, RNA, ligands, and ions. Its accuracy clearly surpassed that of traditional specialized docking tools.

    This shows that the large model paradigm is spilling over from language and multimodality into scientific discovery. Its engineering value lies in sharply compressing experimental screening cycles that used to take months. Its limitations are equally clear: model outputs are hypotheses, not conclusions, and they still need validation in wet-lab experiments. As a result, whether this direction can be industrialized depends on reliable data feedback loops and mechanisms for feeding experimental results back into the models.

    4. China’s Path: From a Model Release Race to Engineering Execution

    In China, large model releases entered a densely packed period in 2023. After Baidu launched a domestically developed conversational large model, several vendors released their own models within just two months. Models with more than 10 billion parameters piled up quickly, forming what the industry called the first wave of the “hundred-model war.”

    This round of competition was marked by a flurry of launches but clear homogenization. It quickly narrowed to engineering capability and real-world deployment. Whoever could provide more stable inference services, lower API costs, and fine-tuning toolchains better suited to specific industries was the one with staying power. On general-purpose cloud platforms (for example, Baidu AI Cloud with its container platform, object storage, logging service, and monitoring and alerting), enterprises can build the full chain from model integration and private deployment to production operations with a relatively low barrier to entry.

    5. Three Questions the Industry Must Answer

    Looking back in 2026, the industry needs to carefully take stock of three questions and answer them.

    5.1 What Can Models Do? Capability Boundaries and Evaluation

    Capabilities have expanded from single-turn question answering to multi-turn reasoning, multimodal understanding, task execution, and scientific prediction. Yet being able to demo a feature does not mean being able to deliver reliably. Engineering teams should build evaluation systems segmented by scenario and difficulty. Quantifiable metrics, such as task completion rate, hallucination rate, and long-context stability, should replace subjective impressions. Teams should also define clearly where a model is and is not suitable.

    5.2 How Should the Industry Be Governed? Regulation and Safety

    As generated content spreads, deepfakes, copyright ownership, and data privacy have become problems that cannot be avoided. A sound governance framework should include:

    • Compliant sources for training data
    • Content moderation of model outputs
    • Provenance marking of generated content
    • Tiered and risk-based regulatory compliance

    Governance is not the opposite of innovation. It is a prerequisite for large-scale deployment.

    5.3 How Should Capital Price It? Commercial Sustainability

    After a hype phase driven by concepts, the market is shifting from “stories about parameters” to “viable business loops.” What truly determines valuation is whether inference costs can keep falling, whether gross margin per API call can turn positive, and whether real willingness to pay for scenarios actually exists. For developers, this means technical choices should prioritize cost control, private deployment, and reproducible evaluation rather than simply chasing leaderboard rankings.

    6. Conclusion

    Between 2022 and 2026, AI underwent a systemic leap:

    • From being able to converse to being able to deliver
    • From an open-source-dominated… from a landscape dominated by closed source to one where open and closed source coexist
    • From single-modality to natively multimodal
    • From predicting the next word to predicting science and code

    For developers and enterprise users, the real opportunity is not chasing every release. It is building sustainable capabilities for evaluation, deployment, and governance that turn general-purpose models into reliable productivity for solving real problems. The technology will keep evolving, but methodology will hold its value for longer.