granite-4.1-3b-mlx-16bit

16-bit MLX conversion of ibm-granite/granite-4.1-3b optimized for Apple Silicon native GPU inference.

Converted by: SirSahOl
Source Model: ibm-granite/granite-4.1-3b
Framework: MLX by Apple
Quantization: 16-bit (Average 16.00 (unquantized bfloat16) bits per weight)
Format: safetensors
License: apache-2.0

---

Model Details

• Architecture: GraniteForCausalLM
• Parameters: 3.0B
• Context Length: 131,072 tokens
• Format: MLX (Apple Silicon native GPU format)
• Quantization: 16-bit (Average 16.00 (unquantized bfloat16) bits per weight)
• Active VRAM Footprint: ~7.1 GB (Minimum recommended: 16 GB Unified Memory)

---

Quick Start

Installation

pip install mlx-lm

Usage

CLI

# Chat interactively
mlx_lm.chat --model SirSahOl/granite-4.1-3b-chat-mlx-16bit

# Generate text
mlx_lm.generate --model SirSahOl/granite-4.1-3b-chat-mlx-16bit --prompt "Write a short poem about artificial intelligence."

Python API (with Chat Template)

from mlx_lm import load, generate

model, tokenizer = load("SirSahOl/granite-4.1-3b-chat-mlx-16bit")

messages = [
    {"role": "user", "content": "Explain quantum superposition in simple terms."}
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)

---

Performance Benchmarks

Apple Silicon Hardware Sizing Matrix

Estimated decoding throughput, time-to-first-token (TTFT), and active unified memory footprint across Apple Silicon tiers:

Projections based on Apple Silicon memory bandwidth saturation for 16-bit weights. Real-world speeds vary by prompt length.

---

Multi-Quantization Comparison

Evaluate your hardware budget and choose the optimal precision:

---

Who Should Use This?

General guidance:

• Use 4-bit if you want to run this model alongside IDEs, browsers, and background development tools.
• Use 8-bit if you have 16GB+ unified memory and require superior reasoning and code accuracy.
• Use 16-bit for research, benchmarking, evaluation, or high-end workstation deployments.

---

Other Quantization Variants

---

LM Studio & Local Inference Setup Guide

To prevent runaway loops and ensure correct conversational turn-taking, configure custom stop tokens in your local inference runtime:

1. <|im_start|>
2. <|im_end|>
3. <|endoftext|>

Prompt Template Formatting

• System Prefix: <|im_start|>system\n
• System Suffix: <|im_end|>\n
• User Prefix: <|im_start|>user\n
• Assistant Suffix: <|im_end|>\n<|im_start|>assistant\n

Ollama Quickstart

FROM SirSahOl/granite-4.1-3b-chat-mlx-16bit
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7

ollama create granite-4.1-3b-chat-mlx-16bit -f Modelfile
ollama run granite-4.1-3b-chat-mlx-16bit

---

Conversion Details

Reproduction

To reproduce this conversion:

pip install mlx-lm==0.31.3
python3 -m mlx_lm.convert --hf-path /root/.cache/huggingface/hub/models--ibm-granite--granite-4.1-3b/snapshots/c0650403e44e78ec0262dab1c90914c65b196c4e --mlx-path output/granite-4.1-3b-mlx-16bit

---

Limitations & Known Issues

• 4-bit group-wise quantization introduces minor precision loss compared to unquantized weights; for deep mathematical derivations or precision-critical reasoning, test the 8-bit or 16-bit variants.
• High context sequences (>32K tokens) require sufficient unified memory headroom; ensure unified memory is not overcommitted.
• This is a weight-only MLX conversion designed specifically for Apple Silicon GPUs (M1/M2/M3/M4 series).

---

License

This model conversion inherits the license of the source model: apache-2.0.

See the original model card for full license details.

---

Changelog

---

Converted with MLX Foundry — a professional pipeline for converting models to Apple MLX format.