---
title: MiMo-V2.5-JANGTQ_2
canonical_url: "https://www.modelscope.ai/models/RealmSky/MiMo-V2.5-JANGTQ_2"
md_url: "https://www.modelscope.ai/models/RealmSky/MiMo-V2.5-JANGTQ_2.md"
repository: RealmSky/MiMo-V2.5-JANGTQ_2
last_updated: 2026-06-13
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - mimo_v2
architectures:
  - MiMoV2ForCausalLM
base_model:
  - XiaomiMiMo/MiMo-V2.5
base_model_relation: finetune
parameters: 21.8B
tensor_type:
  - BF16
  - F16
  - I32
  - U32
  - F32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 171
stars: 0
tags:
  - mimo-v2
  - jangtq
  - mlx
  - osaurus
---

# MiMo-V2.5-JANGTQ_2

> MiMo-V2.5-JANGTQ_2 - An open-source model by RealmSky on ModelScope. MiMo V2.5 JANGTQ 2 for Osaurus

RealmSky/MiMo-V2.5-JANGTQ_2 is a 21.8B-parameter text-generation model on ModelScope. licensed under apache-2.0. derived from XiaomiMiMo/MiMo-V2.5.

- **Repository**: RealmSky/MiMo-V2.5-JANGTQ_2
- **License**: apache-2.0
- **Tasks**: text-generation
- **Parameters**: 21.8B
- **Base model**: XiaomiMiMo/MiMo-V2.5
- **Tags**: mimo-v2, jangtq, mlx, osaurus
- **Downloads**: 171
- **Stars**: 0
- **Last updated**: 2026-06-13

Source: https://www.modelscope.ai/models/RealmSky/MiMo-V2.5-JANGTQ_2

---

![Osaurus](./osaurus-x-banner.png)

# MiMo V2.5 JANGTQ 2 for Osaurus

TurboQuant JANGTQ bundle for MiMo V2.5, prepared for Osaurus/vMLX runtimes.

## Bundle

- Format: JANGTQ
- Profile: JANGTQ_2
- Family: mimo_v2
- Layers: 48
- Routed experts: 256
- TQ layout: prestacked_switch_mlp
- Routed expert bits: 2-bit packed TQ experts
- Runtime sidecar: `jangtq_runtime.safetensors`
- Tokenizer: Qwen2Tokenizer
- Chat template: `chat_template.jinja`
- Verified local size: 79G

## Modalities

Text runtime path is the primary target. Vision and audio weights/tokenizer assets are preserved in the bundle for runtimes that wire MiMo multimodal execution.

## Files

This repo includes the safetensor shard index, quantization metadata, tokenizer files, `generation_config.json`, MiMo configuration code, and the TurboQuant runtime sidecar required by JANGTQ loaders.