OmniDimen-v2.5-4B-Emotion-oQ6
Overview
This is an
oQ6 quantized version of
OmniDimen-v2.5-4B-Emotion, optimized for
Apple Silicon inference via the
MLX framework.
The original model is a fine-tuned variant of Qwen/Qwen3.5-4B, specialized for emotion recognition and emotionally-aware text generation. As a native MLX format, this version also retains the base model's vision-language capability.
| Detail | Value |
|---|
| Base Model | Qwen/Qwen3.5-4B |
| Fine-tune | OmniDimen-v2.5-4B-Emotion |
| Quantization | oQ6 (6-bit mixed-precision) |
| Quantization Tool | oMLX oQ |
| Format | MLX safetensors |
| Size on Disk | ~3.8 GB |
| Context Length | 262,144 tokens (max) |
| License | MIT |
Vision Capability (VLM)
Since Qwen3.5-4B is natively a vision-language model, this MLX quantized version retains full image understanding capability. The vision encoder weights are preserved during oQ quantization.
Quantization Details
This model was quantized using oQ (oMLX's built-in quantization system). Unlike uniform quantization, oQ performs per-layer sensitivity analysis through calibration and dynamically allocates bit precision using a mix of mxfp4, mxfp8, and affine formats where the data indicates they matter most.
Compatibility Notes
During quantization testing on this fine-tuned model, we observed the following:
- oQ6 ✅ — Generates normally. Recommended for use.
- oQ4 ❌ — Model fails to generate output (silent).
- oQ8 ❌ — Model fails to generate output (silent).
- oQ Enhanced (AWQ) ❌ — Reasoning chain breaks; output is incoherent.
These results are specific to this fine-tuned model (based on Qwen3.5-4B). The base Qwen3.5 architecture contains quantization-sensitive layers (e.g., SSM and attention layers) that may behave unpredictably at certain precision levels after fine-tuning. If you plan to quantize other OmniDimen variants, we recommend testing multiple oQ levels before deployment.
Download & Use
With oMLX (Recommended for Mac)
- Install oMLX on your Mac.
- Search and download this model directly from the oMLX admin dashboard.
- Load and chat via the oMLX web dashboard or OpenAI-compatible API at
http://localhost:8000/v1.
With mlx-lm
oQ produces standard MLX safetensors — no custom loader required.
1from mlx_lm import load, generate
2
3model, tokenizer = load("OmniDimen/OmniDimen-v2.5-4B-Emotion-oQ6")
4
5messages = [
6 {"role": "user", "content": "你认为自己是什么?"}
7]
8
9prompt = tokenizer.apply_chat_template(
10 messages, tokenize=False, add_generation_prompt=True
11)
12
13response = generate(
14 model, tokenizer, prompt=prompt, max_tokens=2048
15)
16print(response)
With any MLX-compatible app
This model works with any application that supports MLX safetensors format, including
LM Studio (MLX backend) and other MLX-based tools.
Usage Tips
- Before starting emotional interactions, it is recommended to inform the model of the user's identity (e.g., how OmniDimen should address the user). This helps reduce hallucinations.
- This model is emotion-focused and may not perform as broadly as the base Qwen3.5-4B model.
- Use responsibly with sensitive content.
Acknowledgments
- OmniDimen team for the fine-tuned base model
- oMLX by jundot for the quantization tool and inference server
- Qwen team for the Qwen3.5-4B foundation model
中文说明
概述
本模型是
OmniDimen-v2.5-4B-Emotion 的
oQ6 量化版本,专为
Apple Silicon 设备优化,使用
MLX 框架进行推理。
原始模型基于 Qwen/Qwen3.5-4B 微调,专注于 情感识别与情感感知文本生成。作为原生 MLX 格式,本版本还 保留了基座模型的视觉语言能力。
| 信息 | 内容 |
|---|
| 基座模型 | Qwen/Qwen3.5-4B |
| 微调版本 | OmniDimen-v2.5-4B-Emotion |
| 量化精度 | oQ6(6-bit 混合精度) |
| 量化工具 | oMLX oQ |
| 格式 | MLX safetensors |
| 磁盘大小 | ~3.8 GB |
| 最大上下文长度 | 262,144 tokens |
| 许可证 | MIT |
视觉能力(VLM)
由于 Qwen3.5-4B 原生支持视觉语言,本 MLX 量化版本 完整保留了图像理解能力。oQ 量化过程中视觉编码器权重被保留。
量化说明
本模型使用 oQ(oMLX 内置量化系统)进行量化。与均匀量化不同,oQ 通过校准数据对每一层进行 敏感度分析,动态分配比特精度,混合使用 mxfp4、mxfp8 和 affine 格式,在模型质量和体积之间取得最佳平衡。
量化兼容性记录
在对该微调模型进行量化测试时,我们观察到以下结果:
- oQ6 ✅ — 正常生成,推荐使用。
- oQ4 ❌ — 无法生成输出(静默失败)。
- oQ8 ❌ — 无法生成输出(静默失败)。
- oQ Enhanced(AWQ 模式) ❌ — 推理链断裂,输出不连贯。
以上结果针对该微调模型(基于 Qwen3.5-4B)。Qwen3.5 架构中包含对量化敏感的层(如 SSM 和注意力层),微调后在特定精度下可能表现异常。如需量化其他 OmniDimen 模型变体,建议先测试多个 oQ 精度级别。
下载与使用
使用 oMLX(Mac 推荐)
- 在 Mac 上安装 oMLX。
- 在 oMLX 管理面板中直接搜索并下载本模型。
- 通过 oMLX 网页面板或 OpenAI 兼容 API(
http://localhost:8000/v1)加载并对话。
使用 mlx-lm
oQ 生成的是标准 MLX safetensors 格式,无需自定义加载器。
1from mlx_lm import load, generate
2
3model, tokenizer = load("OmniDimen/OmniDimen-v2.5-4B-Emotion-oQ6")
4
5messages = [
6 {"role": "user", "content": "你认为自己是什么?"}
7]
8
9prompt = tokenizer.apply_chat_template(
10 messages, tokenize=False, add_generation_prompt=True
11)
12
13response = generate(
14 model, tokenizer, prompt=prompt, max_tokens=2048
15)
16print(response)
使用其他 MLX 兼容应用
本模型兼容所有支持 MLX safetensors 格式的应用,包括
LM Studio(MLX 后端)等工具。
使用建议
- 在开始情感互动前,建议先告知模型用户的身份信息(如希望 OmniDimen 如何称呼你),这有助于减少幻觉。
- 本模型 专注于情感对话,在通用任务上的表现可能不如基座 Qwen3.5-4B。
- 请负责任地使用,注意敏感内容。
致谢
- OmniDimen 团队提供微调基座模型
- oMLX(作者 jundot)提供量化工具与推理服务器
- Qwen 团队提供 Qwen3.5-4B 基础模型