---
title: K2-Horizon-MoVA-36B-A4B-8bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/K2-Horizon-MoVA-36B-A4B-8bit"
md_url: "https://www.modelscope.cn/models/mlx-community/K2-Horizon-MoVA-36B-A4B-8bit.md"
repository: mlx-community/K2-Horizon-MoVA-36B-A4B-8bit
last_updated: 2026-09-29
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - k2_horizon
architectures:
  - K2HorizonForCausalLM
base_model:
  - IFM/K2-Horizon-MoVA-36B-A4B
base_model_relation: quantized
parameters: 10.5B
tensor_type:
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 48
stars: 0
tags:
  - mlx
  - quantized
  - moe
  - k2-horizon
  - mova
---

# K2-Horizon-MoVA-36B-A4B-8bit

> K2-Horizon-MoVA-36B-A4B-8bit - mlx-community 在 ModelScope 开源的模型。K2-Horizon-MoVA-36B-A4B (MLX, 8-bit)

mlx-community/K2-Horizon-MoVA-36B-A4B-8bit 是 ModelScope 魔搭社区上的 10.5B 参数text-generation模型，采用 apache-2.0 许可，基于 IFM/K2-Horizon-MoVA-36B-A4B 构建。

- **Repository**: mlx-community/K2-Horizon-MoVA-36B-A4B-8bit
- **License**: apache-2.0
- **Tasks**: text-generation
- **Parameters**: 10.5B
- **Base model**: IFM/K2-Horizon-MoVA-36B-A4B
- **Tags**: mlx, quantized, moe, k2-horizon, mova
- **Downloads**: 48
- **Stars**: 0
- **Last updated**: 2026-09-29

Source: https://www.modelscope.cn/models/mlx-community/K2-Horizon-MoVA-36B-A4B-8bit

---

# K2-Horizon-MoVA-36B-A4B (MLX, 8-bit)

**8-bit** MLX quantization of [IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B), converted from revision `e5c131d`. **8.5 bits/weight effective**, 37 GB on disk. For Apple silicon.

K2-Horizon-MoVA-36B-A4B is IFM's sparse K2-Horizon model: a 36B-parameter Mixture-of-Experts model with 4B active per token and 512K context. Its attention is Mixture-of-Values (MoVA), where the values come from 4 of 64 routed value experts.

## Requirements

mlx-lm doesn't support the `k2_horizon` architecture yet. There's an open request: [mlx-lm#1876](https://github.com/ml-explore/mlx-lm/issues/1876). Until support lands, **this repo ships the MLX model code** ([`k2_horizon.py`](./k2_horizon.py)), which mlx-lm loads through the `model_file` entry in `config.json`. So the released mlx-lm works as-is:

```bash
pip install -U mlx-lm
```

Pass `--trust-remote-code` (or `trust_remote_code=True`). mlx-lm 0.31.3 loads the file without it, but newer versions require it. The file runs on your machine, so read it first if you like.

Once K2-Horizon support lands in a released mlx-lm, this repo will be updated to use it. If something breaks after that, open a discussion here asking for a re-upload.

## How it was quantized

`mlx_lm.convert -q --q-bits 8 --q-group-size 64` (affine). Every linear layer, the routed experts, the MoVA value experts and the embeddings are 8-bit. The MoE and MoVA routers are custom modules and stay in bf16. WikiText-2 perplexity is within 0.04% of bf16, so it's effectively lossless.

## Memory

Peak **40 GB** for a short prompt: fits a **64 GB** Mac under the default Metal working-set cap. The KV cache adds about 192 KB per token (about 24 GB at 128K tokens).

## Conversion check

The MLX implementation was checked against IFM's PyTorch code (`modeling_k2_horizon.py`) in **fp32 on the real weights**, one layer at a time. All 48 layers match to a relative error of 1e-6 or better, and the next-token predictions agree at every position.

Smoke-tested after conversion with released mlx-lm 0.31.3 (`mlx_lm.generate` and `mlx_lm.server`):

- `17 * 23` → `391` with correct reasoning
- "capital of Australia" → `Canberra`
- A multi-turn server conversation that carries context correctly

On a **Mac Studio M4 Max 128GB**: 51.2 tok/s generation, peak 39.9 GB (short prompt).

## Benchmarks (all K2-Horizon MLX variants)

WikiText-2 test perplexity (128 × 512 tokens, lower is better) and generation speed on an M4 Max 128GB, all measured the same way:

| | [bf16](https://huggingface.co/mlx-community/K2-Horizon-MoVA-36B-A4B-bf16) | [8-bit](https://huggingface.co/mlx-community/K2-Horizon-MoVA-36B-A4B-8bit) | [4-bit](https://huggingface.co/mlx-community/K2-Horizon-MoVA-36B-A4B-4bit) |
|---|---|---|---|
| Bits/weight | 16 | 8.5 | 5.6 |
| Disk | 70 GB | 37 GB | 25 GB |
| Peak memory | 75.0 GB | 39.9 GB | 26.5 GB |
| WikiText-2 perplexity | 11.368 | 11.372 (+0.04%) | 11.508 (+1.2%) |
| Generation | 35.3 tok/s | 51.2 tok/s | 59.6 tok/s |

Perplexity is a coarse signal. Test the versions on your own workload before picking one.

## Usage

```bash
mlx_lm.generate --model mlx-community/K2-Horizon-MoVA-36B-A4B-8bit --trust-remote-code --prompt "Explain mixture-of-experts in two sentences." --max-tokens 2048
```

```python
from mlx_lm import load, generate

# Newer mlx-lm versions need trust_remote_code=True; on mlx-lm 0.31.3 use load("mlx-community/K2-Horizon-MoVA-36B-A4B-8bit").
model, tokenizer = load("mlx-community/K2-Horizon-MoVA-36B-A4B-8bit", trust_remote_code=True)
messages = [{"role": "user", "content": "Explain mixture-of-experts in two sentences."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt, max_tokens=2048))
```

`mlx_lm.chat` and `mlx_lm.server` take the same `--trust-remote-code` flag.

The model thinks before it answers, inside `<ifm|think> … </ifm|think>`. Leave room for that in `max_tokens`. The effort level is set with `reasoning_effort` in the chat template: `"high"` (default), `"medium"` or `"low"`, e.g. `apply_chat_template(messages, add_generation_prompt=True, reasoning_effort="low")`.

**Notes:**

- **Server output:** mlx-lm doesn't recognize the `<ifm|think>` tags yet, so `mlx_lm.server` returns the thinking text inside `content`, before `</ifm|think>`, rather than in a separate `reasoning` field. K2-Horizon's tool-call format isn't parsed yet either.
- **Chat template change:** the original template raises an error when an earlier assistant message has no thinking field, which is what OpenAI-style clients send. The template here renders empty thinking for those messages instead. Nothing else was changed.

## License

[Apache-2.0](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B), inherited from the base model. Refer to the [original model card](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B) for architecture, benchmarks and intended use. All credit for the model belongs to IFM.