Instructions to use neureps/warmly-qwen35-2b-enko-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- llama-cpp-python
How to use neureps/warmly-qwen35-2b-enko-gguf with llama-cpp-python:
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="neureps/warmly-qwen35-2b-enko-gguf", filename="Qwen3.5-2B-enko-dIQ4_XS.gguf", )
llm.create_chat_completion( messages = [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] ) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use neureps/warmly-qwen35-2b-enko-gguf with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M # Run inference directly in the terminal: llama cli -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M # Run inference directly in the terminal: llama cli -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Use Docker
docker model run hf.co/neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
- LM Studio
- Jan
- vLLM
How to use neureps/warmly-qwen35-2b-enko-gguf with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "neureps/warmly-qwen35-2b-enko-gguf" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "neureps/warmly-qwen35-2b-enko-gguf", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
- Ollama
How to use neureps/warmly-qwen35-2b-enko-gguf with Ollama:
ollama run hf.co/neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
- Unsloth Studio
How to use neureps/warmly-qwen35-2b-enko-gguf with Unsloth Studio:
Install Unsloth Studio (macOS, Linux, WSL)
curl -fsSL https://unsloth.ai/install.sh | sh # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for neureps/warmly-qwen35-2b-enko-gguf to start chatting
Install Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for neureps/warmly-qwen35-2b-enko-gguf to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for neureps/warmly-qwen35-2b-enko-gguf to start chatting
- Pi
How to use neureps/warmly-qwen35-2b-enko-gguf with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Configure the model in Pi
# Install Pi: npm install -g @mariozechner/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent new
How to use neureps/warmly-qwen35-2b-enko-gguf with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Run Hermes
hermes
- Atomic Chat new
- OpenClaw new
How to use neureps/warmly-qwen35-2b-enko-gguf with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
- Docker Model Runner
How to use neureps/warmly-qwen35-2b-enko-gguf with Docker Model Runner:
docker model run hf.co/neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
- Lemonade
How to use neureps/warmly-qwen35-2b-enko-gguf with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull neureps/warmly-qwen35-2b-enko-gguf:Q4_K_M
Run and chat with the model
lemonade run user.warmly-qwen35-2b-enko-gguf-Q4_K_M
List all available models
lemonade list
warmly-qwen35-2b-enko-gguf — Warmly 프로덕션 모델 (프루닝 2B GGUF)
오프라인 SNS 앱 Warmly(온기)의 프로덕션 배포 리포입니다. EN/KO 어휘 프루닝 마스터 neureps/Qwen3.5-2B-enko를 llama.cpp로 양자화했습니다. Android 앱(내장 llama-server)이 루트의 배포 파일을 직접 다운로드합니다.
버전 이력
| 버전 | 날짜 | 텍스트 본체 | 상태 | 개선 내용 (실측) |
|---|---|---|---|---|
| distill-v1 | 2026-07-15 | Qwen3.5-2B-enko-distill-v1-dIQ4_XS.gguf |
게이트 통과 — 승격 대기 (프로덕션 전환은 별도 결정) | 한국어 자연성 증류 (어댑터 리포 v1 병합). Kanana-1.5-8B 교사 QLoRA — 어색 |
| base | 2026-07-14 (앱 v0.2.15) | Qwen3.5-2B-enko-dIQ4_XS.gguf |
현행 프로덕션 | 어휘 프루닝(−9% 무손실)+도메인 imatrix IQ4_XS — 구 배포(1564MB) 대비 −363MB, 하니스 20/20 |
파일 (루트 — 프로덕션 파일명 변경 금지, 앱이 URL로 고정 참조)
| 파일 | 크기 | 설명 |
|---|---|---|
Qwen3.5-2B-enko-dIQ4_XS.gguf |
975MB | 현행 배포 본체 (base). IQ4_XS + 도메인 imatrix |
Qwen3.5-2B-enko-distill-v1-dIQ4_XS.gguf |
1022MB | 증류 v1 — 게이트 통과, 승격 대기. 증류 병합본 + 도메인 imatrix(재산출) IQ4_XS. heldout PPL 11.85(댓글 도메인 특화로 위키 PPL은 상승 — 품질 지표 아님, 블라인드 승률이 판정) |
Qwen3.5-2B-enko-mmproj-q5_0.gguf |
228MB | 배포 비전 프로젝터(공용). 텍스트 LoRA와 무관해 두 버전이 공유. q8_0 대비 캡션 greedy 구별 불가 |
imatrix-domain-distill-v1.gguf |
1.9MB | distill-v1 병합본에서 재산출한 도메인 imatrix (imatrix는 모델 종속 — base 것 재사용 불가) |
imatrix-domain.gguf |
1.9MB | base용 도메인 imatrix (캘리브레이션 = calib.txt) |
imatrix-general.gguf |
1.9MB | base용 일반 imatrix (캘리브레이션 = calib-general.txt) |
calib.txt |
365KB | 도메인 캘리브레이션 텍스트 — 페르소나 프롬프트 + 앱 학습 JSONL(영어 캡션+한국어 댓글) (양 버전 공용, 재현용) |
calib-general.txt |
2MB | 일반 캘리브레이션 텍스트 — KO/EN 위키+감성 댓글 표본 (재현용) |
현행 배포 구성 합계(base+q5_0) 1201MB — 공식 하니스 20/20, heldout PPL 10.91, v0.2.15부터. distill-v1의 제작 과정·잔존 특성(복잡 캡션 영어 누출 → 후처리+재생성이 흡수, 폴백 0)은 어댑터 리포 카드와 학습 데이터 warmly-distill-data 참조.
양자화 변형 (variants/ — A안 실측 스윕 전체, 2026-07-13)
파일명 규칙: Qwen3.5-2B-enko-{양자화}-{imatrix 종류}.gguf
imatrix-domain= 도메인 캘리브레이션(위imatrix-domain.gguf) /imatrix-general= 위키+댓글(imatrix-general.gguf) /plain= imatrix 없음
| 파일 (variants/) | 크기 | PPL(heldout) | 실측 판정 |
|---|---|---|---|
...-Q4_K_M-plain.gguf |
1.1GB | 10.94 | imatrix 없음 — 한국어 단어 오류("소파/소거"류), 비권장 |
...-Q4_K_M-imatrix-general.gguf |
1.1GB | 10.70 | 동크기 최고 품질 |
...-IQ4_XS-imatrix-general.gguf |
975MB | 10.87 | −12% 크기인데 plain Q4_K_M보다 나음 — 스위트스팟 |
...-Q3_K_M-imatrix-general.gguf |
883MB | 12.10 | +10% 열화, 크기 절박할 때만 |
...-Q4_K_M-imatrix-domain.gguf |
1.1GB | 10.75 | 후처리 통과 19/20, 어색 ~3 |
...-IQ3_M-imatrix-domain.gguf |
845MB | 11.82 | 오역·환각 증가 시작 (usable-but-degraded) |
...-IQ3_XXS-imatrix-domain.gguf |
745MB | 13.33 | 의미 붕괴 — 사용 부적합. 단 후처리 통과 19/20 → 통과율은 저비트 붕괴를 못 잡는다는 실측 증거로 보존 |
...-mmproj-f16.gguf |
638MB | — | 비전 프로젝터 원본 (llama.cpp 변환기는 q8_0까지만 지원) |
...-mmproj-q8_0.gguf |
348MB | — | 구 배포(v1) 페어링 |
...-mmproj-q4_0.gguf |
189MB | — | 자체 재양자화(mmproj_requant.py). 4/5 대등, 애매한 피사체(food)에서 환각 드리프트 — 크기 절박할 때만 |
핵심 교훈: 후처리 통과율은 저비트 열화를 못 잡는다 — 진짜 판별자는 자연성(주관 평가)이며 PPL과 강하게 상관. 2B는 품질을 유지하며 1GB 언더가 불가능(유일한 1GB 언더 IQ3_XXS는 붕괴).
사용법 (llama.cpp)
# 텍스트 (댓글 생성) — enable_thinking:false 필수, 안 끄면 출력이 빔
llama-server -m Qwen3.5-2B-enko-dIQ4_XS.gguf -c 4096
# 비전 (캡션)
llama-mtmd-cli -m Qwen3.5-2B-enko-dIQ4_XS.gguf --mmproj Qwen3.5-2B-enko-mmproj-q5_0.gguf --image photo.jpg -p "Describe..."
관련 리포
- 프루닝 마스터(safetensors): neureps/Qwen3.5-2B-enko
- 구 배포 v1(프루닝 전, 롤백용): neureps/warmly-qwen35-2b-gguf
- 0.8B GGUF: neureps/warmly-qwen35-08b-enko-gguf
- Downloads last month
- 2,044