Instructions to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- llama-cpp-python
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with llama-cpp-python:
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="Sophia-AI/Qwen2.5-32B-Instruct-GGUF", filename="Qwen2.5-32B-Instruct-F16.gguf", )
llm.create_chat_completion( messages = "No input example has been defined for this model task." )
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with llama.cpp:
Install from brew
brew install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama-server -hf Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M # Run inference directly in the terminal: llama-cli -hf Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama-server -hf Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M # Run inference directly in the terminal: llama-cli -hf Sophia-AI/Qwen2.5-32B-Instruct-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 Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf Sophia-AI/Qwen2.5-32B-Instruct-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 Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
Use Docker
docker model run hf.co/Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
- LM Studio
- Jan
- Ollama
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with Ollama:
ollama run hf.co/Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
- Unsloth Studio
How to use Sophia-AI/Qwen2.5-32B-Instruct-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 Sophia-AI/Qwen2.5-32B-Instruct-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 Sophia-AI/Qwen2.5-32B-Instruct-GGUF to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for Sophia-AI/Qwen2.5-32B-Instruct-GGUF to start chatting
- Pi
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama-server -hf Sophia-AI/Qwen2.5-32B-Instruct-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": "Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent new
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama-server -hf Sophia-AI/Qwen2.5-32B-Instruct-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 Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
Run Hermes
hermes
- Docker Model Runner
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with Docker Model Runner:
docker model run hf.co/Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
- Lemonade
How to use Sophia-AI/Qwen2.5-32B-Instruct-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
Run and chat with the model
lemonade run user.Qwen2.5-32B-Instruct-GGUF-Q4_K_M
List all available models
lemonade list
Qwen2.5-32B-Instruct GGUF (imatrix)
Versione GGUF ottimizzata di Qwen2.5-32B-Instruct, quantizzata da Sophia-AI con importance matrix (imatrix) per inferenza efficiente su GPU consumer e server.
Questi sono i modelli ufficiali utilizzati sui nostri progetti.
๐ง Quantizzazione con imatrix: a differenza delle quantizzazioni standard, queste utilizzano una importance matrix generata su un dataset di calibrazione multilingue e multi-task. Questo preserva i pesi piรน critici del modello durante la compressione, con miglioramenti significativi soprattutto su Q2_K e Q3_K_M.
๐ฆ File disponibili
| File | Quant | Dimensione | Score | Consigliato per |
|---|---|---|---|---|
Qwen2.5-32B-Instruct-F16.gguf |
F16 | ~65 GB | 8/8 โ | Massima qualitร / Fine-tuning |
Qwen2.5-32B-Instruct-Q8_0.gguf |
Q8_0 | ~34 GB | 8/8 โ | GPU con 48+ GB VRAM |
Qwen2.5-32B-Instruct-Q6_K.gguf |
Q6_K | ~26 GB | 8/8 โ | GPU con 32+ GB VRAM |
Qwen2.5-32B-Instruct-Q5_K_M.gguf |
Q5_K_M | ~23 GB | 8/8 โ | GPU con 24+ GB VRAM |
Qwen2.5-32B-Instruct-Q4_K_M.gguf |
Q4_K_M | ~19 GB | 8/8 โ | โญ Best balance โ GPU 24 GB |
Qwen2.5-32B-Instruct-Q3_K_M.gguf |
Q3_K_M | ~15 GB | 8/8 โ | ๐ฅ GPU 16 GB / Split GPU+CPU |
Qwen2.5-32B-Instruct-Q2_K.gguf |
Q2_K | ~12 GB | 8/8 โ | GPU 16 GB / CPU con 32+ GB RAM |
Tutti i quant passano 8/8 test funzionali โ tool calling, JSON output, codice, italiano โ senza degradazione rispetto a F16.
๐ง Perchรฉ imatrix?
La quantizzazione standard tratta tutti i pesi del modello allo stesso modo: li comprime uniformemente. Questo funziona bene per Q8 e Q6, ma a livelli aggressivi (Q3, Q2) i pesi critici per tool calling, JSON strutturato e ragionamento multilingue vengono compressi troppo.
L'importance matrix (imatrix) risolve questo problema:
- Eseguiamo un forward pass del modello F16 su un dataset di calibrazione rappresentativo
- Misuriamo quali pesi si attivano maggiormente (= sono piรน importanti)
- Durante la quantizzazione, i pesi importanti vengono compressi meno, quelli meno importanti di piรน
Il risultato: Q2_K con imatrix mantiene capacitร che senza imatrix sarebbero perse.
Dataset di calibrazione
Il nostro dataset รจ stato costruito specificamente per riflettere i casi d'uso reali di Sophia Metal:
| Sezione | Contenuto | Scopo |
|---|---|---|
| ๐ฎ๐น Italiano (~30%) | Articoli Wikipedia italiani | Preservare capacitร multilingue |
| ๐ฌ๐ง Inglese (~30%) | WikiText (train + valid + test) | Baseline linguistica |
| ๐ป Codice (~20%) | Python, JavaScript/Next.js, SQL, bash, YAML | Preservare generazione codice |
| ๐ง Chat + Tools (~20%) | Conversazioni Qwen2.5 con tool calling, JSON output, multi-turn | Preservare tool calling e output strutturato |
La sezione Chat + Tools รจ quella piรน critica: include esempi reali di <tool_call>, risposte JSON strutturate, tool definitions nei system prompt, parallel tool calls e conversazioni multi-turn. Senza questi pattern nel dataset di calibrazione, i pesi responsabili della generazione di {"name": "get_weather", "arguments": ...} verrebbero identificati come poco importanti e compressi troppo.
โ Test di qualitร
Tutti i quant sono stati testati con una suite automatizzata che verifica le capacitร core per l'uso in produzione:
Test F16 Q8_0 Q6_K Q5_K_M Q4_K_M Q3_K_M Q2_K
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
JSON output semplice โ
โ
โ
โ
โ
โ
โ
Tool call - singolo โ
โ
โ
โ
โ
โ
โ
Tool call - multiplo โ
โ
โ
โ
โ
โ
โ
Tool call - scelta tool โ
โ
โ
โ
โ
โ
โ
Tool follow-up โ
โ
โ
โ
โ
โ
โ
Italiano - coerenza โ
โ
โ
โ
โ
โ
โ
Codice Python โ
โ
โ
โ
โ
โ
โ
JSON strutturato complesso โ
โ
โ
โ
โ
โ
โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Score 8/8 8/8 8/8 8/8 8/8 8/8 8/8
I test verificano:
- Tool calling nativo (singolo, multiplo, selezione del tool corretto)
- Tool follow-up (interpretazione della risposta JSON del tool)
- JSON output (semplice e strutturato complesso con nested arrays)
- Italiano (coerenza linguistica)
- Codice (generazione Python con logica corretta)
๐ฏ Quale scegliere?
๐ฅ๏ธ Desktop / Server GPU
| La tua VRAM | Modello consigliato | Note |
|---|---|---|
| 80 GB+ (A100/H100) | Q8_0 o Q6_K | Massima qualitร |
| 48 GB (A6000/RTX 6000 Ada) | Q8_0 | Full offload |
| 32 GB (V100) | Q6_K o Q5_K_M | Ottimo compromesso |
| 24 GB (RTX 3090/4090) | Q4_K_M โญ | Sweet spot |
| 16 GB (RTX 4080/4070 Ti) | Q3_K_M o Q2_K | Partial offload consigliato |
๐ฅ Multi-GPU / Split GPU+CPU
Per chi ha GPU con VRAM limitata ma vuole sfruttare la potenza di un 32B:
| Setup | Modello | Note |
|---|---|---|
| 2x RTX 3090/4090 (48 GB tot) | Q5_K_M o Q6_K | Qualitร eccellente |
| RTX 4090 + CPU offload | Q4_K_M โญ | Buon compromesso velocitร /qualitร |
| RTX 4080 16 GB + CPU offload | Q3_K_M | Funziona bene |
๐ก Usa
n_gpu_layersper controllare quante layers caricare su GPU. Le layers rimanenti vengono eseguite su CPU.
๐ฅ Sophia Metal / Edge Computing (8GB unified memory)
โ ๏ธ Il modello 32B รจ troppo grande per dispositivi edge con 8GB di memoria unificata. Per edge computing su Sophia Metal, consigliamo Qwen3-4B-Instruct-2507-GGUF con Q3_K_M.
Per dispositivi edge con memoria unificata elevata (32+ GB, es. Apple Silicon M2 Pro/Max/Ultra):
| Setup | Modello | Note |
|---|---|---|
| 64 GB unified | Q4_K_M โญ | Best balance su Apple Silicon |
| 32 GB unified | Q3_K_M o Q2_K | Funziona bene grazie a imatrix |
๐ป CPU Only
| RAM disponibile | Modello |
|---|---|
| 64 GB+ | Q4_K_M |
| 48 GB | Q3_K_M |
| 32 GB | Q2_K |
๐ฅ Download
# โญ Consigliato per la maggior parte degli utenti (RTX 3090/4090)
huggingface-cli download Sophia-AI/Qwen2.5-32B-Instruct-GGUF \
Qwen2.5-32B-Instruct-Q4_K_M.gguf
# ๐ฅ Per GPU 16 GB / Split GPU+CPU
huggingface-cli download Sophia-AI/Qwen2.5-32B-Instruct-GGUF \
Qwen2.5-32B-Instruct-Q3_K_M.gguf
# Massima qualitร (48+ GB VRAM)
huggingface-cli download Sophia-AI/Qwen2.5-32B-Instruct-GGUF \
Qwen2.5-32B-Instruct-Q6_K.gguf
๐ Utilizzo
llama-cpp-python
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="Sophia-AI/Qwen2.5-32B-Instruct-GGUF",
filename="*Q4_K_M.gguf",
n_ctx=4096,
n_gpu_layers=-1, # -1 = tutte le layers su GPU
)
response = llm.create_chat_completion(
messages=[
{"role": "user", "content": "Ciao, come stai?"}
]
)
print(response["choices"][0]["message"]["content"])
Tool calling
response = llm.create_chat_completion(
messages=[
{"role": "user", "content": "What's the weather in Rome?"}
],
tools=[{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get weather for a city",
"parameters": {
"type": "object",
"properties": {
"city": {"type": "string"}
},
"required": ["city"]
}
}
}]
)
llama.cpp CLI
./llama-cli -m Qwen2.5-32B-Instruct-Q4_K_M.gguf \
-p "<|im_start|>user\nCiao!<|im_end|>\n<|im_start|>assistant\n" \
-n 256 -ngl 99
Ollama
ollama run hf.co/Sophia-AI/Qwen2.5-32B-Instruct-GGUF:Q4_K_M
๐ Specifiche modello
| Parametro | Valore |
|---|---|
| Parametri | 32.5B |
| Context length | 131,072 tokens |
| Architettura | Qwen2.5 |
| Vocab size | 152,064 |
| Layers | 64 |
| Attention heads | 40 |
| KV heads | 8 (GQA) |
๐ง Dettagli quantizzazione
| Parametro | Valore |
|---|---|
| Sorgente | Qwen2.5-32B-Instruct F16 (convertito da HuggingFace safetensors) |
| Tool | llama.cpp llama-quantize con --imatrix |
| imatrix | Generata con llama-imatrix, chunk size 512 |
| Calibration dataset | ~10MB, multilingue (IT/EN), codice (Python/JS/SQL/bash), chat Qwen2.5 con tool calling |
| GPU per imatrix | NVIDIA (full GPU offload) |
๐ Links
- Sophia Metal: metal.2sophia.ai
- Sophia AI: 2sophia.ai
- Modello base: Qwen/Qwen2.5-32B-Instruct
๐ Licenza
Apache 2.0 โ Stesso del modello base Qwen2.5.
๐ Crediti
- Downloads last month
- 41
2-bit
3-bit
4-bit
5-bit
6-bit
8-bit
16-bit