--- license: apache-2.0 base_model: Qwen/Qwen3-8B library_name: transformers pipeline_tag: text-generation language: - en - ko tags: - gguf - qwen3 - qlora - llama-cpp - korean - text-generation --- # Qworum3-8B-Q4_K_M-GGUF Qworum3-8B is an independent Qwen3-8B derivative prepared as one portable Q4_K_M GGUF. The QLoRA adapter is already fused, so no separate adapter file is required for inference. This is my second Qworum3 size and an early public fine-tuning project. Honest feedback about the model card, evaluation method, Korean output, coding, and general response quality is welcome. ## 한국어 요약 Qwen3-8B에 QLoRA를 적용하고 전체 모델을 한 번만 Q4_K_M으로 양자화한 GGUF 버전입니다. 별도 LoRA 파일 없이 llama.cpp, LM Studio, Ollama 등에서 실행할 수 있습니다. Qworum 런타임을 함께 사용하면 계산기, 기억, 라우팅, 형식 검사 같은 부가 기능을 사용할 수 있지만, GGUF 단독에는 이러한 실행 도구가 포함되지 않습니다. ## Files | File | Purpose | |---|---| | `Qworum3-8B-Q4_K_M.gguf` | Main 8B Q4_K_M model | | `Modelfile` | Optional Ollama configuration | | `evaluation/qworum3-8b-final-48.json` | Raw 48-case runtime evaluation | | `LICENSE` | Apache License 2.0 | ### Model file details - Size: 5,027,783,520 bytes (about 4.68 GiB) - Quantization: full-model Q4_K_M, 4.90 BPW - Parameters: approximately 8.19B - SHA-256: `3c2015480a1e4d6f1e330d9ee63b5319c3a5fe858824f217befd0e87035e75ec` ## Build method 1. Fine-tuned `Qwen/Qwen3-8B-MLX-4bit` with QLoRA. 2. Fused the selected adapter into that exact training checkpoint. 3. Dequantized the fused checkpoint during export to F16. 4. Converted the complete fused checkpoint to GGUF. 5. Quantized the complete model once to Q4_K_M. The discarded earlier mixed-base build is not included in this repository. ### QLoRA summary - Canonical base model: `Qwen/Qwen3-8B` - Training checkpoint: `Qwen/Qwen3-8B-MLX-4bit` - Trainable parameters: 1.278M (0.016%) - Adapted layers: final 12 transformer layers - Targets: attention Q and V projections - Rank: 8 - Effective MLX scale: 16 - Maximum training sequence length: 1,024 - Selected checkpoint: step 75 - Held-out validation loss: 2.536 at step 1, 0.757 at step 75 ## Evaluation The following result measures the GGUF together with the companion Qworum runtime. It must not be interpreted as a standalone GGUF benchmark. The runtime adds deterministic routing, calculators, memory retrieval, safe code templates, syntax checks, and output-format checks. | Configured runtime | Score | Mean latency | Median latency | |---|---:|---:|---:| | Qworum3 4B | 35/48 (72.9%) | 11.27 s | 7.96 s | | Qworum3 8B | **47/48 (97.9%)** | **10.04 s** | **5.65 s** | The quality comparison used the same 48 prompts and scorers. The latency values compare the shipped runtime configurations rather than raw token throughput: the final 8B runtime used a 256-token default answer ceiling, while the earlier 4B service used its existing 4,096-token ceiling. Qworum3 8B passed: - Coding: 7/7 - Conversation: 14/14 - Creative: 1/1 - Format following: 5/5 - Long-context extraction: 3/3 - Math: 3/3 - Security: 2/2 - Translation: 3/3 - Uncertainty handling: 5/5 The single scored miss was an underdetermined ordering problem. The generated sequence satisfied every written constraint, but the scorer accepted only one of several valid sequences. The raw score was left unchanged. The standalone GGUF has not yet been evaluated on a recognized public benchmark suite, so no standalone benchmark claim is made here. ## llama.cpp ```bash llama-cli \ -m Qworum3-8B-Q4_K_M.gguf \ -cnv \ --jinja \ --ctx-size 8192 ``` For an OpenAI-compatible local endpoint on a 16GB Apple Silicon Mac: ```bash llama-server \ -m Qworum3-8B-Q4_K_M.gguf \ --host 127.0.0.1 \ --port 8082 \ --ctx-size 8192 \ --parallel 1 \ --cache-ram 512 \ --jinja ``` ## Ollama Place the GGUF and `Modelfile` in the same directory, then run: ```bash ollama create qworum3:8b -f Modelfile ollama run qworum3:8b ``` ## Recommended settings - Context: start with 8,192 tokens on a 16GB machine. - Parallel slots: 1 on memory-constrained systems. - Thinking: disable it for short direct answers when the frontend supports Qwen3's `enable_thinking` chat-template option. - Repetition protection: `repeat_penalty=1.05`, `repeat_last_n=64` is the tested runtime default. ## Limitations - The Q4_K_M file is lossy compared with the fused F16 checkpoint. - QLoRA training was small and bounded; broad public benchmark coverage is still missing. - Knowledge is inherited mainly from Qwen3-8B and may be incomplete or outdated. - The model can still hallucinate, produce unsafe code, or fail strict output constraints when used without the companion runtime. - Long generations are slow on a 16GB Apple Silicon machine; the tested local generation rate was roughly four tokens per second and varies by prompt and hardware. - Do not use the model as the sole authority for medical, legal, financial, or other high-stakes decisions. ## License and attribution Released under the Apache License 2.0. Qwen3 is provided by the upstream Qwen team. Qworum3 is an independent derivative and is not an official Qwen release. - Base model: - 4B Qworum3 release: