This is an unofficial oQ8e conversion of Thinking Machines Lab’s Inkling 2 Mini for MLX on Apple Silicon. It accepts text and images and generates text. The four target shards use 18.12 GB. The original DSpark draft adds 0.64 GB. The complete checkpoint is about 18.78 GB, or 7.0% smaller than the source snapshot (decimal sizes).
Inkling 2 Mini is a mixture-of-experts model. Its middle 12 of 24 physical text layers run three times per token. This conversion keeps that execution order, the vision encoder, the tokenizer, and the chat template. No training or fine-tuning was done.
The source already used NVFP4 and MXFP8 for the routed experts. oMLX kept those formats and quantized selected other weights. This is not an all-8-bit checkpoint.
| Saved weight format | Projection groups or modules | Treatment |
|---|---|---|
| MLX affine 8-bit, group size 64 | 58 | oQ8e applied importance weights to 57. The token embedding has no importance entry. |
| NVFP4, group size 16 | 40 | Retained from the source. |
| MXFP8, group size 32 | 8 | Retained from the source. |
| BF16 DSpark draft | Separate file | mtp.safetensors is a byte-for-byte copy. |
Other tensors follow oMLX’s precision rules. BF16 was the requested output dtype. The quantizer used its built-in oqe_code_multilingual calibration set: 128 samples of 512 tokens in batches of 16. It calibrated the native source checkpoint, without a lower-precision proxy. All 57 collected importance-matrix entries were used, with no missing entries or shape mismatches reported for eligible weights.
Use oMLX 0.8.1 on Apple Silicon. This is the release used to create and check the model. Download all files together and keep mtp.safetensors next to config.json if you want to use the bundled draft.
hf download a-gordo/Inkling-2-Mini-oQ8e --local-dir ./models/Inkling-2-Mini-oQ8e
Select ./models as a model directory in oMLX and load Inkling-2-Mini-oQ8e. The weights use the oMLX Inkling 2 runtime. This is not a GGUF file or a Transformers/ModelOpt checkpoint. Use the source model card for its intended uses, prompt guidance, and limitations.
Exact file sizes, checksums, and build settings are in quantization_info.json.
ed40e7d.13fd0c9.The original model and this conversion use the Apache License 2.0. See the original model card, including its safety and limitations sections and its link to the Model Acceptable Use Policy.
model-00003-of-00004.safetensors 5.2 GBmodel-00002-of-00004.safetensors 5.2 GBmodel-00001-of-00004.safetensors 5.1 GBmodel-00004-of-00004.safetensors 2.6 GBthinkingmachines/Inkling-2-Mini