Open-Sourcing a 9B Visual Reasoning Model: A Technical Breakdown of the Multimodal Reasoning Training Framework

    This article provides a technical analysis of a recently open-sourced 9B-parameter visual reasoning model and its multimodal training framework. It explores the shift towards chain-of-thought reasoning in vision-language models and details a two-stage training approach, featuring large-scale pretraining and Reinforcement Learning with Curriculum Sampling (RLCS). Furthermore, it highlights the three foundational pillars of the framework—data quality, precise reward systems, and efficient RL strategies—and discusses why a 9B model can outperform larger 70B+ models, offering actionable insights for developers building multimodal applications.

    Background: Why Multimodal Models Need to “Learn to Think First”

    The evolution of vision-language models (VLMs) can be roughly divided into three stages. Early models primarily focused on “perception,” tackling tasks like image captioning and visual question answering. Mid-stage models developed image-text alignment and cross-modal retrieval capabilities to handle more complex inputs. The current trend, however, is to equip models with the ability to “think clearly.” When faced with tasks such as scientific calculations, chart analysis, long-document comprehension, or GUI interactions, models must perform chain-of-thought (CoT) reasoning before arriving at a conclusion.

    This shift is driven by the fact that real-world multimodal problems are rarely solvable through single-step perception. For example, when analyzing a diagram for a physics problem, a model must not only recognize the physical quantities and symbols in the image but also perform multi-step derivations using formulas. If a model merely maps an image directly to an answer, the error rate escalates rapidly as problem complexity increases. Consequently, next-generation models treat reasoning capability as the core training objective rather than a byproduct.

    The newly open-sourced model is designed precisely with this philosophy in mind. It utilizes a CoT output format: presenting the complete reasoning process first, followed by the final answer. This design offers two immediate benefits. First, the reasoning process can be manually reviewed and evaluated, enhancing interpretability. Second, the intermediate reasoning steps provide richer optimization signals for reinforcement learning (RL), allowing the model’s capabilities to be systematically improved.

    Training Framework Breakdown: Pretraining Foundation and RL Enhancement

    The overall training pipeline can be summarized as “two stages, one core”:

    Stage 1: Large-Scale Pretraining for Foundational Visual Capabilities. The model is first pretrained on massive datasets of images, text, and videos to establish a robust multimodal perception foundation. This covers basic tasks such as image comprehension, video temporal understanding, document and chart recognition, and visual grounding. This stage determines the model’s “breadth of knowledge” and “visual acuity”—essentially, what types of content it can understand.

    Stage 2: Reinforcement Learning with Curriculum Sampling (RLCS). This is the core innovation of the framework. Traditional RL training typically samples from a fixed task distribution, which often leads to two issues: tasks that are too difficult yield sparse reward signals, causing the model to learn nothing, while tasks that are too easy result in reward saturation and vanishing gradients. The curriculum sampling approach dynamically sorts and sequences tasks by difficulty, progressing from easy to hard, much like human learning:

    1. # Schematic logic for curriculum sampling strategy (pseudo-code)
    2. def sample_task(model, task_pool):
    3. # Dynamically adjust sampling probabilities for tasks of varying difficulty based on real-time model performance
    4. accuracy = evaluate(model, task_pool)
    5. if accuracy < 0.4: # Poor performance, fall back to easy tasks for consolidation
    6. return task_pool.easy.sample()
    7. elif accuracy > 0.85: # Excellent performance, advance to harder tasks
    8. return task_pool.hard.sample()
    9. else:
    10. return task_pool.medium.sample()

    Through this dynamic adjustment, the model consistently trains within a difficulty zone that is challenging but achievable. This keeps the reward signals effective and significantly boosts learning efficiency.

    Broad Task Domain Coverage. The report notes that this framework achieves comprehensive improvements across multiple dimensions, including STEM problem-solving, video understanding, content recognition, code generation, visual grounding, GUI agents, and long-document comprehension. This is particularly noteworthy: previous efforts often applied RL to single tasks, leading to a “seesaw effect” where improving mathematical capabilities might degrade general conversational abilities. The multi-task RLCS design demonstrates that carefully crafted curricula and rewards can prevent such capability trade-offs.

    Three Key Elements: Data, Reward, and Strategy

    The technical report emphasizes three pillars for building a general multimodal reasoning model, which serve as a critical methodological summary for the industry:

    1. High-Quality Data as the Foundation. Constructing multimodal reasoning data is far more difficult than text-only data. Detail annotations in images, numerical verification in charts, and temporal event alignment in videos all carry the risk that any noise will directly pollute the training signal. Reasoning data construction must ensure logical consistency among the “problem,” the “reasoning chain,” and the “answer,” otherwise the model will learn incorrect reasoning patterns. In practice, this typically requires a multi-layered filtering pipeline combining rule-based validation, cross-model verification, and manual spot-checks.

    2. Precise Reward Systems as the Steering Wheel. The essence of reinforcement learning is “adjusting behavior based on rewards.” If the reward function is even slightly biased, the model will exploit it (reward hacking). In multimodal scenarios, reward design must cover multiple dimensions: the correctness of the final answer, the validity of the reasoning steps, format compliance, and the presence of hallucinations. For instance, visual grounding tasks can calculate rewards based on the Intersection over Union (IoU) between predicted and ground-truth bounding boxes; mathematical tasks can be scored based on exact numerical matches; and open-ended tasks require verifier models to assign scores, with the verifier models themselves undergoing adversarial validation.

    3. Efficient RL Strategies as the Amplifier. The computational cost of multimodal RL training is significantly higher than text-only scenarios, as every sample involves visual encoding and long-sequence reasoning. Common engineering optimizations include decoupling asynchronous sampling from the training pipeline, controlling the length of reasoning steps to reduce sequence overhead, and leveraging the sample efficiency gains inherent to curriculum sampling. For small-parameter models, training efficiency directly dictates whether they can be sufficiently refined within a limited budget.

    What the Small Model’s Comeback Tells Us

    A model with approximately 9 billion parameters outperforming 72B-level models across multiple tasks is backed by clear technical logic:

    • Inference-time computation compensates for parameter scale. Chain-of-thought outputs allow the model to “calculate a few extra steps” during inference, effectively trading inference time for model capacity. This trade-off is particularly advantageous for smaller models.
    • Higher marginal returns from RL fine-tuning. While large models hold clear advantages in data and compute during the pretraining phase, the quality of reward signals and the efficiency of training strategies become the dominant factors during the RL phase. Smaller models can actually be “trained more thoroughly.”
    • Structural advantages from task focus. Curriculum design precisely directs the model’s capabilities toward the skill domains covered by evaluations, avoiding the “average distribution” of parameters.

    This phenomenon is highly relevant for enterprise technology selection. In privatized deployments or cost-sensitive scenarios, small-to-medium models that have undergone rigorous reasoning training often offer better cost-performance ratios than under-optimized massive models. They also provide lower inference latency and more accessible deployment thresholds. Developers can validate performance locally using the open-source weights, and then scale deployments and integrate monitoring systems via cloud model services or container platforms.

    Practical Insights for Developers

    Based on the analysis above, enterprises can draw on the following experiences when building multimodal applications:

    1. Prioritize evaluating reasoning chains, not just answers. When building evaluation sets, incorporate intermediate reasoning steps into the scoring dimensions. This helps identify systematic flaws in the model earlier.
    2. Ensure evaluation sets cover the full task spectrum. Establish separate benchmarks for perception, reasoning, grounding, document, and GUI comprehension capabilities. Avoid making selection decisions based on a single metric.
    3. Focus on emerging scenarios like long documents and GUIs. Interface agents and long-context document comprehension are becoming critical capabilities for automated office scenarios and are worth including in technical roadmaps.
    4. Build your own data and reward loop. If fine-tuning is required in a vertical domain, refer to the framework philosophy of “curriculum sampling + multi-dimensional reward.” Use business data to construct task pools of varying difficulties, paired with composite reward functions based on both rules and model verification.

    Conclusion

    The value of this work lies not only in open-sourcing a high-performance model but also in validating a clear technical route: reasoning-centric, grounded by high-quality data, directed by precise rewards, and amplified by curriculum-based RL strategies. For technical teams focused on multimodal directions, this framework philosophy offers universally applicable insights. Whether utilizing open-source models or cloud-based multimodal capabilities, understanding the trinity training methodology of “data-reward-strategy” will help teams make smarter technology selections and implementation decisions.