Jev-Style-0.8B-Decision-v3-mlx-8bit

8-bit MLX conversion of chaoliangUNSW/Jev-Style-0.8B-Decision-v3 optimized for Apple Silicon native GPU inference.

Converted by: SirSahOl
Source Model: chaoliangUNSW/Jev-Style-0.8B-Decision-v3
Framework: MLX by Apple
Quantization: 8-bit (Average 8.25 bits per weight)
Format: safetensors
License: apache-2.0

---

Model Details

• Architecture: Qwen3_5ForCausalLM
• Parameters: 0.8B
• Context Length: 262,144 tokens
• Format: MLX (Apple Silicon native GPU format)
• Quantization: 8-bit (Average 8.25 bits per weight)
• Active VRAM Footprint: ~990 MB (Minimum recommended: 8 GB Unified Memory)

---

Quick Start

Installation

pip install mlx-lm

Usage

CLI

# Chat interactively
mlx_lm.chat --model SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-8bit

# Generate text
mlx_lm.generate --model SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-8bit --prompt "Write a short poem about artificial intelligence."

Python API (with Chat Template)

from mlx_lm import load, generate

model, tokenizer = load("SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-8bit")

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

Measured Benchmarks (Apple M1)

Metric	8-bit
Tokens/sec	52.23
TTFT	19.15 ms
Peak Memory	919.3 MB

Benchmarked on Apple M1 with 8GB unified memory. Average over 5 runs with 256 max tokens.

---

Multi-Quantization Comparison

Evaluate your hardware budget and choose the optimal precision:

Variant	Disk Size	VRAM Footprint	Target Apple Silicon Hardware	Key Advantage
4-bit MLX	~450 MB	~490 MB	M1 / M2 / M3 / M4 (8GB+ Unified Memory)	Ultra-compact footprint, maximum generation speed; negligible memory pressure.
8-bit MLX (This Repository)	~855 MB	~900 MB	M1 / M2 / M3 / M4 (8GB+ Unified Memory)	Near-lossless precision with an extremely lightweight footprint.
16-bit MLX	~1620 MB	~1690 MB	M1 / M2 / M3 / M4 (8GB+ Unified Memory)	Full unquantized precision; zero perplexity penalty for reference evaluation.

---

Who Should Use This?

Your Hardware	Recommended Quantization
M1/M2/M3/M4 (8GB – 16GB)	4-bit — Best balance of speed, low memory, and multitasking capability
M1/M2/M3/M4 Pro/Max (18GB – 36GB)	8-bit — Higher quality reasoning with comfortable memory headroom
M1/M2/M3/M4 Max/Ultra (36GB – 192GB)	16-bit — Unquantized full precision, zero quality degradation

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

Variant	Link
4-bit	SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-4bit
8-bit	SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-8bit
16-bit	SirSahOl/Jev-Style-0.8B-Decision-v3-mlx-16bit

---

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/Jev-Style-0.8B-Decision-v3-mlx-8bit
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7

ollama create jev-style-0.8b-decision-v3-mlx-8bit -f Modelfile
ollama run jev-style-0.8b-decision-v3-mlx-8bit

---

Conversion Details

Property	Value
Source Model	chaoliangUNSW/Jev-Style-0.8B-Decision-v3
Quantization	8-bit
mlx-lm Version	0.31.3
Conversion Time	4.76s
Output Size	782.1 MB
Date	2026-10-13T00:25:12.271253+00:00

Reproduction

To reproduce this conversion:

pip install mlx-lm==0.31.3
python3 -m mlx_lm.convert --hf-path /Users/z4/.cache/huggingface/hub/models--chaoliangUNSW--Jev-Style-0.8B-Decision-v3/snapshots/b023d1f9c7858fbf01504577a3bfc349ea5c7385 --mlx-path output/b023d1f9c7858fbf01504577a3bfc349ea5c7385-mlx-8bit -q --q-bits 8

---

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

Version	Date	Changes
v1.0	2026-10-13	Initial conversion

---

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