Recommended starter~5.24 GB
Q4_K_M
Compact and practical first choice. Start here unless you already know your machine has enough RAM or VRAM for larger variants.
pb_qwen3_5_9b_dora_v7_stable_q4_k_m.gguf
PowerBASIC model library
Additional AI Model Downloads
PowerBASIC-focused Qwen 9B DoRA GGUF model variants for local AI runtimes such as LM Studio, Ollama, llama.cpp-compatible tools, KoboldCpp, text-generation-webui, and other GGUF-capable local inference environments.
These downloads are optional companions for users who want local model behavior tuned toward PowerBASIC-oriented code assistance. Choose the quantization level that matches your available RAM/VRAM.
The GGUF files are hosted in the shared download area outside AISPR and can be used in LM Studio, Ollama imports, llama.cpp-compatible runners, KoboldCpp, text-generation-webui, and similar local model tools.
Start simple
Start with Q4_K_M unless you know your machine has enough memory for larger variants.
Raise quality
Use Q6_K when you want stronger output quality and can spend more memory.
Reference use
Use Q8_0 or F16/BF16 only for capable hardware, archival, conversion, or reference workflows.
Hugging Face
Repository and model cards
All current model files and release metadata are published in one repository, including sizes, checksums, and direct download links.
Available files
Qwen 9B PowerBASIC GGUF variants
Hosted on Hugging Face: https://huggingface.co/Theogott/pb-qwen3_5-9b-powerbasic-ggufs
Lightweight~4.98 GB
Q4_K_S
Even more compact with good quality. Ideal for CPU offloading or when memory is tight.
pb_qwen3_5_9b_dora_v7_stable_q4_k_s.gguf
Smallest variant~4.95 GB
Q4_0
The smallest and fastest variant. Best choice for pure CPU inference or very limited memory.
pb_qwen3_5_9b_dora_v7_stable_q4_0.gguf
Higher quality~6.85 GB
Q6_K
Use this when you can spend more memory for stronger answers and more stable code-oriented behavior.
pb_qwen3_5_9b_dora_v7_stable_q6_k.gguf
High precision~8.87 GB
Q8_0
For capable hardware where quality matters more than download size and runtime memory footprint.
pb_qwen3_5_9b_dora_v7_stable_q8_0.gguf
Maximum quality~16.7 GB
F16
For demanding workflows with high precision, e.g. code analysis, conversion, and detailed testing.
pb_qwen3_5_9b_dora_v7_stable_f16.gguf
Maximum quality GPU~16.7 GB
BF16
Brain-float16 variant for maximum GPU performance while maintaining high quality.
pb_qwen3_5_9b_dora_v7_stable_bf16.gguf
Now available
SPR Models
Hosted on Hugging Face: https://huggingface.co/Theogott/spr-qwen3_5-9b-dora-vramsafe-gguf
SPR GGUF~5.2 GB
Q4_K_M VRAM-safe
VRAM-safe variant with same token-space structure plus a safer memory footprint.
spr_qwen3_5_9b_dora_vramsafe_q4_k_m.gguf
SPR GGUF~6.9 GB
Q6_K VRAM-safe
Higher precision while keeping memory pressure manageable.
spr_qwen3_5_9b_dora_vramsafe_q6_k.gguf
SPR GGUF~8.9 GB
Q8_0 VRAM-safe
For stronger output quality in local runtimes with enough memory available.
spr_qwen3_5_9b_dora_vramsafe_q8_0.gguf
SPR GGUF~16.7 GB
BF16 VRAM-safe
Higher precision variant for advanced inference and conversion workflows.
spr_qwen3_5_9b_dora_vramsafe_bf16.gguf