---

license: apache-2.0

base_model: Qwen/Qwen3-4B

pipeline_tag: text-generation

library_name: gguf

language:

- en

- ko

tags:

- qwen3

- gguf

- qlora

- conversational

- llama-cpp

- ollama

model-index:

- name: Qworum3-4B-Q4_K_M-GGUF

results: []

---

Qworum3-4B-Q4_K_M-GGUF

Qworum3-4B is an experimental 4.02B-parameter local assistant derived from

Qwen/Qwen3-4B. A QLoRA adapter was

trained on a mixture of conversation, reasoning, coding, tool-use, and

memory-related examples. The step-50 checkpoint was selected using held-out

validation loss, merged into the base model, and quantized to GGUF Q4_K_M.

This repository contains one portable weight file:

| File | Quantization | Size | SHA-256 |

|---|---|---:|---|

| Qworum3-4B-Q4_K_M.gguf | Q4_K_M, about 4.95 BPW | about 2.3 GiB | ac37687c87d4ee4e882fad3b6925ccbe409a912f116c5b087485618df588504e |

What is included

The download contains the merged and quantized model weights. When loaded directly

in llama.cpp, Ollama, or LM Studio, it works as a normal standalone language model.

Long-term memory, web search, calculators, clocks, code execution, test-based code

repair, output checking, and the assisted long-form writing workflow are separate

companion software. They are not embedded in the GGUF file.

This distinction matters when reading the evaluation below: the 1,000-question

result measures the GGUF deployment with that companion runtime enabled, not the

weight file in isolation.

Run with llama.cpp

llama-server \
  -m Qworum3-4B-Q4_K_M.gguf \
  --host 127.0.0.1 \
  --port 8081 \
  --ctx-size 8192 \
  --jinja

Thinking is normally disabled in the Qworum runtime for lower latency. With an

OpenAI-compatible client, pass chat_template_kwargs: {"enable_thinking": false}

when supported.

Run with Ollama

Place the included Modelfile beside the GGUF, then run:

ollama create qworum3:4b -f Modelfile
ollama run qworum3:4b

1,000-question GGUF evaluation

On July 12, 2026, a frozen local test set contained 1,000 unique prompts: 100

prompts in each of ten fields. The Qworum3 GGUF was evaluated through the Qworum

companion runtime. The original Qwen3-4B comparison used deterministic greedy

generation.

| System | Correct | Accuracy | Mean latency |

|---|---:|---:|---:|

| Qworum3 GGUF Q4_K_M + companion runtime | 952/1000 | 95.2% | 5.051 s |

| Original Qwen3-4B 4-bit baseline | 659/1000 | 65.9% | 7.8 s |

Results by field

| Field | Qworum3 GGUF | Original Qwen3-4B |

|---|---:|---:|

| Basic arithmetic | 100/100 | 64/100 |

| Algebra | 100/100 | 72/100 |

| Geometry and measurement | 100/100 | 68/100 |

| Statistics and probability | 100/100 | 58/100 |

| Number theory and sequences | 100/100 | 48/100 |

| Logic and computing | 98/100 | 58/100 |

| Science and humanities | 93/100 | 75/100 |

| Time calculations and instruction following | 100/100 | 64/100 |

| Executable coding | 72/100 | 72/100 |

| Constrained creative writing | 89/100 | 80/100 |

The benchmark was created for this project and is not a standard public

leaderboard. Its score includes help from deterministic calculation, tool routing,

format checks, code execution, and targeted output repair. It therefore describes

the complete local deployment and must not be read as a standalone model-weight

score. The original frozen results are reported without retroactively changing

failed answers.

Known limitations

• Standalone coding quality is not consistently better than the Qwen3-4B base.
• Long-form fiction remains weak. In a 30-paragraph stress test assisted by the

companion runtime, generation took about 42 minutes on an Apple M4 and produced

only 2,280 Korean characters; most paragraphs required runtime repair.

• Small-model repetition, grammatical degradation, and multilingual script mixing

can still occur. The companion software checks some of these failures, but the

raw GGUF does not.

• The model inherits the knowledge, language coverage, biases, and hallucination

risks of its Qwen3-4B base.

• The GGUF metadata advertises a 65,536-token training context, but the tested local

server configuration used an 8,192-token context. Long-context quality has not

been established.

• This is an experimental research release, not a safety-critical or high-stakes

decision system.

Training and conversion summary

• Base: Qwen/Qwen3-4B
• Method: QLoRA, universal mixed-domain adapter
• Selected checkpoint: step 50
• Fusion: adapter fused into the 4B base
• Conversion: fused checkpoint to GGUF F16
• Published quantization: Q4_K_M
• Tested engine: llama.cpp on Apple Silicon/Metal

License and attribution

Qworum3-4B follows the Apache License 2.0 lineage of Qwen3-4B. See LICENSE

and NOTICE. Qwen and Qwen3 are projects of the upstream Qwen team; this derivative

release is not an official Qwen release and is not endorsed by the upstream team.