Instructions to use mii-llm/nev-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use mii-llm/nev-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 mii-llm/nev-gguf:F16
# Run inference directly in the terminal:
llama cli -hf mii-llm/nev-gguf:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf mii-llm/nev-gguf:F16
# Run inference directly in the terminal:
llama cli -hf mii-llm/nev-gguf:F16

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf mii-llm/nev-gguf:F16
# Run inference directly in the terminal:
./build/bin/llama cli -hf mii-llm/nev-gguf:F16

Use Docker

docker model run hf.co/mii-llm/nev-gguf:F16

Llama new
LM Studio
Jan
Ollama
How to use mii-llm/nev-gguf with Ollama:
```
ollama run hf.co/mii-llm/nev-gguf:F16
```
Unsloth Desktop

How to use mii-llm/nev-gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf mii-llm/nev-gguf:F16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "mii-llm/nev-gguf:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use mii-llm/nev-gguf with Docker Model Runner:
```
docker model run hf.co/mii-llm/nev-gguf:F16
```

Lemonade

How to use mii-llm/nev-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull mii-llm/nev-gguf:F16

Run and chat with the model

lemonade run user.nev-gguf-F16

List all available models

lemonade list

Hermes Agent

How to use mii-llm/nev-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 mii-llm/nev-gguf:F16

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 mii-llm/nev-gguf:F16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use mii-llm/nev-gguf with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf mii-llm/nev-gguf:F16

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 "mii-llm/nev-gguf:F16" \
  --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"

Nev GGUF — standalone bilingual decisions

Nev v3 is a small English/Italian decision model. Given state, instructions and candidate options, it returns categorical, boolean or ordinal decisions with probabilities. It does not generate conversational text. This public repository contains its decoder in F16 and Q8_0, plus the original FP32 pointer head, tokenizer and calibration configuration. Nesso3 is not required.

Install and use Nev alone on CPU

The tested runtime is an Apple M3 Pro with four CPU threads, Metal/CUDA disabled. Requires Python 3.12 or 3.13, Git, a C/C++ compiler (Xcode command-line tools on Mac), and network access to Hugging Face and GitHub.

git clone https://github.com/mii-llm/nev.git
cd nev/experiments/cpu_gguf
./setup.sh --model nev --precision f16
# Public download; no Hugging Face login is required.
.venv/bin/python decide.py
.venv/bin/python test_standalone.py --model nev --precision f16

The selected download contains only Nev artifacts. The shared Python environment includes conversion/reference dependencies, but inference uses a CPU llama.cpp worker and NumPy. No Nesso3 weights are downloaded or loaded.

The default example routes an Italian damaged-parcel/refund request to refund. To provide your own request, save this as request.json:

{
  "state": "Il pacco è arrivato rotto e voglio un rimborso.",
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Seleziona il tipo di richiesta.",
      "criteria": {
        "refund": "Rimborso o merce danneggiata",
        "account": "Accesso all'account",
        "info": "Informazioni sui prodotti"
      }
    }
  }
}

.venv/bin/python decide.py --request request.json
# Smaller/faster alternative:
./setup.sh --model nev --precision q8
.venv/bin/python decide.py --precision q8 --request request.json

The output contains answers.route.choice, probabilities, token usage and latency. You can supply English state/instructions/criteria in the same format. A high probability is a score over the supplied alternatives, not a calibrated guarantee of correctness.

Python, from the same experiment folder:

import os
os.environ["USE_TORCH"] = "0"
os.environ["HF_HUB_OFFLINE"] = "1"
from cpu_backend import NevGGUF

model = NevGGUF(precision="f16", threads=4)
try:
    result = model.evaluate(
        state="My parcel arrived broken. I want a refund.",
        questions={"route": {
            "type": "choice",
            "instructions": "Select the appropriate department.",
            "criteria": {"refund": "Damaged goods and refunds",
                         "account": "Account access",
                         "info": "Product information"}
        }},
    )
    print(result["answers"])
finally:
    model.close()

Why Nev needs an adapter

This is a decision backbone GGUF, not a standalone chat model. A generic chat application loading the GGUF alone will not reproduce Nev decisions. The supplied adapter reads final-normalized hidden states at option/readout positions, applies the original FP32 pointer head and calibration temperature, and renders each question as an isolated causal row. Keep models/nev/v3/config.json, decision_head.safetensors and tokenizer/ together. No labels are generated as text.

Conversion results

On 20 diagnostic requests containing 24 questions, F16 preserved 24/24 decisions relative to original FP32 CPU inference; maximum probability difference was 0.00541. Q8 preserved 23/24, with maximum difference 0.06086. The changed case was an Italian dependent notes request. Q8 probabilities have not been recalibrated. F16 is the default for closer fidelity, not bit-identical FP32 arithmetic. This small panel does not measure broad task accuracy; the original Nev release documents wider evaluations and limitations.

One F16 file is approximately 843 MiB; Q8 approximately 451 MiB. These are file sizes, not RAM requirements. CPU Mac inference is tested; mobile deployment is not validated.

Files and provenance

gguf/nev-v3-f16.gguf, gguf/nev-v3-q8.gguf: decision decoder only.
models/nev/v3/: original head, decision configuration and tokenizer.
ARTIFACTS.json: exact inference-file sizes and SHA-256 hashes.
PROVENANCE.json: source model and conversion revisions.

Source: mii-llm/nev, folder v3/, revision c5c84121263b423f8846811a24790ab96e66cee3. Conversion uses unmodified llama.cpp 37d6d653e49aa73c947cc9631a5d499869a21cca, with a separate CPU bridge. No retraining or recalibration occurred during conversion. Apache-2.0; source notices included.

For the optional Nesso3 + Nev agent on CPU, see the GitHub combination guide. Nesso3's independent generative GGUF release is mii-llm/nesso3-0.4B-agentic-gguf.