M-26
Collection
Business focused Models • 1 item • Updated
Maverick 2
M-26 (Mid)
M-26(Mid) is an experimental small language model developed as part of my ongoing Maverick project. With fewer than 200 million parameters, it represents my current most capable model in practical use and is an important step in my journey toward building more capable, general-purpose AI systems.
Overview
M-26(Mid) is designed to explore general language modeling, question answering, reasoning, and business-oriented tasks. While it has significant limitations, its development has helped me better understand the challenges of training small language models and the differences between benchmark performance and practical capability.
One of the most interesting findings during development came from comparing M-26(Mid) with M-26(High). Despite M-26(High) achieving a higher score on an internal benchmark, M-26(Mid) demonstrated a better understanding of certain datasets and was able to answer questions about their contents that M-26(High) struggled with.
This highlighted an important lesson in my research: a higher benchmark score does not necessarily translate into better performance on every real-world task. Understanding what a model learns, how it generalizes, and where it fails is just as important as measuring its training or evaluation scores.
Business-Oriented Training
A distinguishing aspect of M-26(Mid) is its emphasis on business-related data and tasks.
My long-term goal is to develop Maverick models capable of making decisions, solving problems, planning, and eventually operating businesses with increasing levels of autonomy. I see business management as a demanding test of AI capabilities because it requires more than generating coherent text. It involves reasoning, evaluating trade-offs, using information effectively, and making decisions toward long-term objectives.
M-26(Mid) is an early step toward exploring these capabilities. It is not currently capable of independently running a business, and its business-oriented training should be understood as an experimental direction rather than a demonstrated level of autonomy.
Development Philosophy
The Maverick project is built through experimentation, evaluation, and iterative improvement. Each model provides insights that inform the development of subsequent versions.
M-26(Mid) has been particularly valuable in helping me investigate how dataset composition, training methods, model architecture, and evaluation strategies affect a model's actual capabilities.
Although M-26(High) is a newer model in the same development line, M-26(Mid) remains an important reference point for comparing architectural changes and measuring progress.
Limitations
M-26(Mid) is a small, experimental language model and should not be expected to perform at the level of larger, instruction-tuned language models.
Its limitations include inconsistent conversational ability, imperfect reasoning, and unreliable performance on tasks that require multiple steps or robust generalization. Its benchmark results also do not guarantee strong performance on unfamiliar tasks.
The model is primarily a research and development project intended to explore what can be achieved with a relatively small parameter budget.
What's Next?
M-26(Mid) is one stage in the broader development of Maverick. Future iterations aim to improve reasoning, conversational competence, problem-solving, instruction following, and practical decision-making.
Ultimately, the goal is to move beyond models that simply generate text toward AI systems that can understand objectives, develop plans, use tools, evaluate outcomes, and work toward complex goals.
M-26(Mid) is not the finished product. It is a milestone in the process of building Maverick into a more capable general-purpose AI system.
Thanks for checking out the project!