---
license: apache-2.0
arxiv: 2608.23283
language:
  - en
  - zh
base_model: apodex/Apodex-1.1-mini
base_model_relation: quantized
pipeline_tag: text-generation
library_name: mlx
tags:
  - mlx
  - mlx-lm
  - 4-bit
  - qwen3.5
  - moe
  - apodex
---

# Apodex-1.1-mini MLX

MLX conversion of [apodex/Apodex-1.1-mini](https://huggingface.co/apodex/Apodex-1.1-mini), a 35.95B-parameter Qwen3.5 MoE model designed for research, data, files, code, and tool-driven work.

This release is optimized for text generation with `mlx-lm` on Apple silicon.

## Upstream benchmarks

Both figures are reproduced from the upstream model card. They report evaluations performed with [FrontierAgent](https://github.com/ApodexAI/FrontierAgent), Apodex's open-source agent runtime and evaluation suite, using its ReAct and Agent Team workflows. They do not measure this MLX conversion.

![Apodex-1.1 benchmark results](./assets/apodex-1.1-benchmarks.png)

*Upstream Apodex-1.1 evaluation results using FrontierAgent's ReAct and Agent Team workflows.*

![Apodex-1.1-mini benchmark results](./assets/apodex-1.1-mini-benchmarks.png)

*Upstream Apodex-1.1-mini evaluation results using FrontierAgent. With the Agent Team workflow, the model scored 27.7 on APEX-Agent, 50.2 on FrontierFinance, and 51.7 on FrontierScience-Research.*

## Model details

| Item | Value |
|---|---|
| Format | MLX safetensors |
| Quantization | Affine 4-bit, group size 64 |
| Effective bits per weight | 4.503 |
| Model size | 19.5 GB |
| Non-quantized dtype | BF16 |
| Architecture | `Qwen3_5MoeForConditionalGeneration` |

The MoE router and shared-expert gate weights remain at 8-bit; other eligible weights use affine 4-bit quantization.

> This is a text-only MLX release. Image input is not supported. Current official `mlx-lm` Qwen3.5 support filters the native `mtp.*` weights, so native MTP is not included. Standard generation is unaffected.

> Hugging Face may show approximately 5B parameters because packed MLX `uint32` tensors are counted as stored elements. The source model contains 35.95B parameters.

## Upstream compatibility

The tokenizer configuration and chat template are synchronized with [upstream revision `62583b4`](https://huggingface.co/apodex/Apodex-1.1-mini/commit/62583b464707ad58848bc2ab066747d8d5b02533).

## Usage

```bash
pip install -U mlx-lm

mlx_lm.generate \
  --model abenzerps/Apodex-1.1-mini-MLX \
  --prompt "Explain why reproducible builds matter." \
  --max-tokens 512 --temp 0.7
```

The 19.5 GB weights require additional unified memory for the runtime and KV cache, especially at long context lengths. To disable thinking, pass `--chat-template-config '{"enable_thinking": false}'`.

## Checksums

SHA-256 hashes for every shard and supporting file are provided in `SHA256SUMS`.

```bash
sha256sum -c SHA256SUMS
```

## License

The source model was released by [Apodex AI](https://huggingface.co/apodex) under the [Apache License 2.0](https://huggingface.co/apodex/Apodex-1.1-mini/blob/27479c1e48453eb8172e03aaeb54eab8ce9419d6/LICENSE). The same license and attribution requirements apply to this conversion.