Exploring Performance and Quality in Advanced Language Models: A Comparative Analysis of RTX 5090 Variants
In recent developments within the AI community, PrismML has introduced a new iteration of their Bonsai model. This version is a heavily quantized adaptation of Qwen 3.8, boasting over 98% top-1 accuracy compared to FP16 benchmarks, with model sizes ranging from approximately 6GB in Q1 to about 8GB in Q2. This release has sparked interest among researchers and enthusiasts eager to evaluate its capabilities relative to other models within similar memory footprints.
To assess the practical performance and output quality of these models, a series of tests were conducted on an RTX 5090 graphics card. The evaluation centered around processing a complex voxel Japanese pagoda scene scripted in Three.js, using an identical prompt for each model to ensure a fair comparison. Notably, despite Bonsai utilizing a higher token budget (primarily dedicated to reasoning), the tests aimed to reflect each model’s autonomy and reasoning efficiency.
Experimental Setup
The testing environment was configured as follows:
| Parameter | Specification |
|———————–|——————————————————————————-|
| Hardware | NVIDIA RTX 5090 with 32GB VRAM |
| Runtime Environment | Llama-server featuring PrismML’s llama.cpp fork (release: prism-b10683-d8f26ee) |
| Software Flags | -ngl 999 -fa on -c 262144 -np 1 --jinja |
| Token Context Length | 262,144 tokens |
| Sampling Parameters | Temperature: 1.0; Top-p: 0.95; Top-k: 20; Min-p: 0 |
| Reasoning Mode | Enabled across all models; Bonsai set to high reasoning level by default |
| Output Limit | None |
| Testing Approach | Single run per model |
| Prompt | A detailed 770-word description of a voxel Japanese pagoda scene |
Models Evaluated and Results
| Model | Quantization | Size | Output Tokens | Approximate Runtime | Throughput |
|————————|————–|———|—————-|———————|——————|
| Ternary Bonsai 2 27B | PQ2_0 | 7.2 GB | 106,396 | ~17.5 minutes | ~101 tokens/sec |
| Qwen 3.5 9B | Q6_K | 7.5 GB | 12,105 | ~70 seconds | ~168 tokens/sec |
| Gemma 4 12B | Q8_0 | 12.7 GB | 9,159 | ~95 seconds | ~96 tokens/sec |
Qualitative Observations
The results reveal notable differences in output quality. The Bonsai 2 27B model demonstrates superior scene detail and coherence, effectively capturing the complexity of the voxel pagoda. Gemma 4 12B offers respectable outputs, though the images appear more washed out and less detailed. Conversely, Qwen 3.5 9B produced some inconsistent and broken outputs, indicating limitations at this size within the evaluated scenario.
Implications and Conclusions
This evaluation underscores the impressive capabilities of the PrismML Bonsai 2 27B model. Despite its quantized nature and modest memory requirements, it delivers high-quality, detailed scene renderings with substantial reasoning effort, highlighting its suitability for resource-constrained environments. Remarkably, such performance suggests that this model can be effectively run on hardware as accessible as a NVIDIA GeForce RTX 3060, enabling broader community participation in advanced AI applications.
The findings are particularly encouraging for local AI development, demonstrating that with optimized quantization and careful configuration, high-quality AI generation is attainable without extreme hardware demands.
Additional Information
The experiments were conducted using the atomic.chat platform, leveraging PrismML’s backend tailored for such models. As a co-founder of PrismML, I welcome feedback and discussions surrounding these results and the evolving landscape of efficient, powerful language models.
Final Thoughts
Advancements like PrismML’s Bonsai 2 27B exemplify the promising trajectory of AI model compression, balancing size, performance, and quality. Continued exploration and community engagement will be vital in unlocking the full potential of these innovations for diverse applications.
