Akicou/ViSWE-GGUF Model Card

Model Details

• Name: Akicou/ViSWE-GGUF
• Base Model: Akicou/ViSWE (derived from Skywork/Skywork-SWE-32B)
• Merged Models: Skywork/Skywork-SWE-32B, TIGER-Lab/VisCoder2-32B
• Merge Method: Arcee Fusion (MergeKit) - selectively merges important elements via dynamic thresholds
• Original Dtype: bfloat16
• Architecture: Qwen2.5-Coder-32B-Instruct based, 33B params
• Tensor Type: BF16 (original); Quantized in GGUF format
• Quantization Format: GGUF (various quantization levels for efficient inference)

Description

This repository contains GGUF quantized versions of the Akicou/ViSWE model, which is a merged model combining Skywork-SWE-32B (code-agent for software-engineering tasks like bug fixing on GitHub) and VisCoder2-32B (executable visualization code generation across 12 languages with self-debugging). These quantizations enable efficient inference on hardware with lower memory requirements, such as consumer-grade GPUs or CPUs, while maintaining high performance for the model's intended tasks.

The original model can be found at: Akicou/ViSWE.

Available GGUF Quants

The following GGUF quantized files are available in this repository:

• ViSWE.Q2_K.gguf
• ViSWE.Q3_K.gguf
• ViSWE.Q3_K_M.gguf
• ViSWE.Q3_K_S.gguf
• ViSWE.Q4_0.gguf
• ViSWE.Q4_K.gguf
• ViSWE.Q4_K_M.gguf
• ViSWE.Q4_K_S.gguf
• ViSWE.Q5_0.gguf
• ViSWE.Q5_K.gguf
• ViSWE.Q5_K_M.gguf
• ViSWE.Q5_K_S.gguf
• ViSWE.Q6_K.gguf
• ViSWE.Q8_0.gguf
• ViSWE.f16.gguf
• ViSWE.fp16.gguf

These quants range from lower-bit (e.g., Q2_K for minimal memory usage) to higher-bit (e.g., Q8_0 or f16 for better accuracy). Choose based on your hardware constraints and desired balance between speed, memory, and precision.

Intended Uses

• Software-engineering tasks (e.g., bug fixing, feature implementation)
• Visualization code generation and rendering
• Multi-language support for executable visuals
• Iterative self-debugging
• Efficient inference on resource-constrained devices using GGUF-compatible runners like llama.cpp

Limitations

• Requires multiple GPUs or sufficient CPU/RAM for inference on larger quants (32B params, 32K context)
• Performance varies by quantization level, task complexity, repository, and language; lower quants may introduce minor accuracy losses
• No explicit safety or bias mitigations
• Quantized models are optimized for inference only and may not support fine-tuning

Training Data

• Skywork-SWE: 8,209 SWE trajectories
• VisCoder2: VisCode-Multi-679K dataset for visualizations

Citation

• Skywork-SWE: Zeng et al. (2025), arXiv:2506.19290
• VisCoder2: Ni et al. (2025), arXiv:2510.23642