INTERPRETIC's picture
Upload quantized weights and model card
3c348d7 verified
|
Raw History Blame Contribute Delete
2.38 kB
metadata
license: apache-2.0
license_link: https://www.apache.org/licenses/LICENSE-2.0
library_name: mlx
pipeline_tag: image-text-to-text
base_model:
  - thinkingmachines/Inkling-2-Mini
tags:
  - inkling
  - moe
  - mlx

Inkling-2-Mini-MLX-4bit

Inkling 2 Mini is a multimodal looped mixture-of-experts model from Thinking Machines Lab: 26.4B parameters, 3.4B active (5.4B effective with three passes through the looped middle layers), text and image input, up to 1M tokens of context. This repository is an unmodified quantization of the official checkpoint (revision 1fb4134).

MLX build for Apple Silicon (MLX, oMLX, LM Studio). Runs on a Mac with 32 GB of unified memory or more.

Other formats from INTERPRETIC AI: GGUF Q4_K_M 路 GGUF IQ3_XXS 路 MLX 4-bit

Files

Four model-0000N-of-00004.safetensors shards (17 GB in total), the configuration and the tokenizer. The vision tower is included in bfloat16. The routed experts keep the original NVFP4 quantization. Attention, embedding and output weights are 4-bit (group size 64). SHA-256 checksums are listed in SHA256SUMS.

Usage

Inkling 2 Mini support in mlx-lm is in ml-explore/mlx-lm#2086 and not yet in a release.

pip install -U "git+https://github.com/ml-explore/mlx-lm.git@refs/pull/2086/head"
mlx_lm.generate --model InterpreticAI/Inkling-2-Mini-MLX-4bit --prompt "Hello" --max-tokens 512

Tested with text input. Image input through mlx-vlm has not been tested.

Notes

  • Inkling 2 Mini is a thinking model: it writes its reasoning before the answer. Use a chat client or server that applies the model's chat template.
  • Quality checks here are spot checks (English, Chinese, code and one image), not a full benchmark.
  • Support for this architecture is new in all runtimes. Use a recent build.

License

Apache 2.0, as the original model. See the model card of Thinking Machines Lab and its acceptable use policy.

INTERPRETIC AI 路 interpretic.ai