Granite-4.2-3B GGUF (llama.cpp)

Quantized GGUF files for Granite-4.2-3B, optimized for deployment with llama.cpp. These are the exact quantized versions of the Granite-4.2-3B model, packaged as GGUF format for efficient inference.

Model Overview

Granite-4.2-3B is IBM's compact reasoning model in the Granite 4.2 family. Despite its small parameter count of 3B, it delivers strong performance on reasoning-intensive tasks by leveraging built-in <think>...</think> chain-of-thought reasoning. It supports flexible thinking modes — full thinking (default), non-thinking, and low-effort — allowing users to balance depth vs. latency on a per-query basis.

Key Capabilities

• Built-in Reasoning: Native chain-of-thought that significantly improves performance on math, coding, and complex multi-step problems
• Flexible Thinking Modes: Seamlessly switch between full thinking, non-thinking, and low-effort modes within a single model
• Reasoning-Augmented Tool Calling: The model reasons about which tools to invoke and why, producing more accurate function calls
• 512K Context Window: Supports long documents, multi-turn conversations, and complex agentic workflows
• Apache 2.0 Licensed: Fully open for commercial and research use

Available Quantizations

All files are cut from the BF16 source (6.82 GB). Each quant is independently uploaded and deleted immediately after successful upload to minimize peak disk usage.

Model Architecture

Granite-4.2-3B is built on a decoder-only dense transformer architecture:

• Attention: Grouped Query Attention (GQA) with 40 attention heads and 8 KV heads
• Position Embedding: Rotary Position Embedding (RoPE) with θ = 10,000,000
• Feed-Forward: MLP with SwiGLU activation (hidden size 8192)
• Normalization: RMSNorm (ε = 1e-5)
• Embeddings: Separate input/output embeddings (not tied)
• Precision: bfloat16 (source)

Inference

Generation Parameters

Important: Use temperature=1.0 and top_p=0.95 across all tasks and serving backends, including general chat, reasoning, and tool calling.

Thinking Modes

Serving with llama.cpp

Basic Usage

# Download a quantization (recommended: Q4_K_M)
huggingface-cli download NANI-Nithin/granite-4.2-3b-GGUF granite-4.2-3b-Q4_K_M.gguf --local-dir .

# Run inference
llama-cli -m granite-4.2-3b-Q4_K_M.gguf -p "Hello, how are you?"

Or use the Hugging Face Hub integration:

# Serve with llama-server
llama-server -hf NANI-Nithin/granite-4.2-3b-GGUF:Q4_K_M

# Run with model path
llama-cli -hf NANI-Nithin/granite-4.2-3b-GGUF:Q4_K_M -p "Hello"

API Usage

# Start the server
llama-server -hf NANI-Nithin/granite-4.2-3b-GGUF:Q4_K_M

# Query with curl
curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "granite-4.2-3b-Q4_K_M",
    "messages": [{"role": "user", "content": "Explain quantum computing in simple terms"}],
    "temperature": 1.0,
    "top_p": 0.95,
    "max_tokens": 8192
  }'

Quick Start Example

# Download the model
huggingface-cli download NANI-Nithin/granite-4.2-3b-GGUF granite-4.2-3b-Q4_K_M.gguf

# Run inference
./llama-cli -m granite-4.2-3b-Q4_K_M.gguf -p "What is the Riemann hypothesis?"

Technical Details

• Source Model: ibm-granite/granite-4.2-3b
• Author: Granite Team, IBM
• License: Apache 2.0
• Languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese
• Architecture: GraniteForCausalLM (decoder-only dense transformer)
• Context Length: 131K tokens (natively supports 128K, extended to 512K)
• Parameters: 3B
• Created: August 25, 2026

Usage Notes

1. Disk Space: Download one quantization at a time. Each file ranges from 0.65 GB (Q1_0) to 6.82 GB (BF16 source).
2. Memory Requirements: Varies by quantization. Q4_K_M requires ~2 GB VRAM, Q8_0 requires ~4 GB VRAM.
3. Performance: All files are optimized for llama.cpp with appropriate kernel selections for each quantization.
4. Thinking Mode: Enable enable_thinking=True to get chain-of-thought reasoning. Set to False for faster, direct answers.

Model Card Information

This GGUF repo contains quantized versions of Granite-4.2-3B, including evaluation results and comprehensive serving instructions. The quantization was performed using llama.cpp's quantization pipeline, preserving the model's reasoning capabilities while reducing size for deployment.

For the full source model documentation, including detailed training methodology, evaluation benchmarks, and advanced usage, refer to the source repo: ibm-granite/granite-4.2-3b.

Note: These files are optimized for llama.cpp and are not compatible with vLLM, SGLang, or the Transformers library in their current format. For those frameworks, use the source model ibm-granite/granite-4.2-3b.