Built with Inkling (Thinking Machines Lab).
MLX (Apple Silicon) conversion of thinkingmachines/Inkling, quantized to 4-bit (affine group quant, group size 64).
Code / loader: github.com/PipeNetwork/inkling-mlx
Inkling is a 975B-total / 41B-active sparse-MoE, natively multimodal model (text + image/video + audio → text). This is the full multimodal conversion: all three towers (text backbone, HMLP vision, dMel audio) are ported; the multi-token-prediction head is dropped (inference-irrelevant).
| Variant | Size | Notes |
|---|---|---|
| 8bit | ~937 GB | near-lossless |
| 6bit | ~717 GB | high quality |
| 4bit | ~490 GB | balanced default |
MLX supports FP4 modes and Thinking Machines ships an Inkling-NVFP4 checkpoint — so for the record, we benchmarked round-trip reconstruction error (‖W − Ŵ‖ / ‖W‖ vs bf16) on real Inkling expert weights:
| Scheme | bits/weight | reconstruction error |
|---|---|---|
| affine int4 (group 64) | 4.50 | ~9.1% |
| nvfp4 (group 16) | 4.50 | ~10.2% |
| mxfp4 (group 32) | 4.25 | ~12.3% |
Affine int4 is the most faithful: it is asymmetric (per-group scale and zero-point, 16 uniform levels), which centers on Inkling's near-Gaussian expert weights better than symmetric FP4's fixed non-uniform levels. FP4's real payoff is heavy-tailed activations and native Blackwell FP4 tensor cores — neither helps weight fidelity on Apple Silicon, where MLX would dequantize FP4 anyway. So these builds use affine int4.
inkling_mlx loaderThe inkling_mm_model architecture is not in stock mlx-lm / mlx-vlm, so this
repo bundles a minimal, numerically-validated MLX implementation under inkling_mlx/.
pip install mlx mlx-lm transformers
from inkling_mlx.load import load
from inkling_mlx.generate import greedy_generate
from transformers import AutoTokenizer
model, config = load("/path/to/this/repo")
tok = AutoTokenizer.from_pretrained("/path/to/this/repo", trust_remote_code=True)
ids = tok("The capital of France is")["input_ids"]
print(tok.decode(greedy_generate(model, config, ids, max_new_tokens=64)))
Needs an Apple-Silicon Mac with enough unified memory to hold the weights (≈ the size above).
InklingProcessor — image patchify/normalize, audio log-mel→dMel,
validated ~1e-7 vs the reference) are included. Pass images/audio via the processor.Conversion is streaming (tensor-by-tensor; the ~1.9 TB bf16 model never fully loads into RAM) and was validated with fp32 numerical parity against transformers PR #47347. License: Apache-2.0 (inherits the base model).
model-00080-of-00097.safetensors 6.5 GBmodel-00040-of-00097.safetensors 6.4 GBmodel-00042-of-00097.safetensors 6.0 GBmodel-00047-of-00097.safetensors 5.8 GBmodel-00002-of-00097.safetensors 5.8 GBmodel-00064-of-00097.safetensors 5.8 GBmodel-00071-of-00097.safetensors 5.7 GBmodel-00052-of-00097.safetensors 5.7 GBmodel-00049-of-00097.safetensors 5.7 GBmodel-00045-of-00097.safetensors 5.7 GBmodel-00092-of-00097.safetensors 5.7 GBmodel-00033-of-00097.safetensors 5.6 GBmodel-00023-of-00097.safetensors 5.6 GBmodel-00020-of-00097.safetensors 5.6 GBmodel-00010-of-00097.safetensors 5.6 GBmodel-00029-of-00097.safetensors 5.6 GBmodel-00069-of-00097.safetensors 5.6 GBmodel-00018-of-00097.safetensors 5.6 GBmodel-00094-of-00097.safetensors 5.6 GBmodel-00084-of-00097.safetensors 5.6 GBmodel-00026-of-00097.safetensors 5.6 GBmodel-00063-of-00097.safetensors 5.6 GBmodel-00082-of-00097.safetensors 5.6 GBmodel-00016-of-00097.safetensors 5.6 GBmodel-00039-of-00097.safetensors 5.6 GBmodel-00038-of-00097.safetensors 5.6 GBmodel-00028-of-00097.safetensors 5.6 GBmodel-00014-of-00097.safetensors 5.6 GBmodel-00089-of-00097.safetensors 5.6 GBmodel-00024-of-00097.safetensors 5.6 GBmodel-00088-of-00097.safetensors 5.6 GBmodel-00019-of-00097.safetensors 5.6 GBmodel-00037-of-00097.safetensors 5.5 GBmodel-00044-of-00097.safetensors 5.5 GBmodel-00009-of-00097.safetensors 5.5 GBmodel-00078-of-00097.safetensors 5.5 GBmodel-00005-of-00097.safetensors 5.5 GBmodel-00093-of-00097.safetensors 5.5 GBmodel-00032-of-00097.safetensors 5.5 GBmodel-00073-of-00097.safetensors 5.5 GBmodel-00004-of-00097.safetensors 5.5 GBmodel-00079-of-00097.safetensors 5.5 GBmodel-00075-of-00097.safetensors 5.5 GBmodel-00087-of-00097.safetensors 5.5 GBmodel-00068-of-00097.safetensors 5.5 GBmodel-00048-of-00097.safetensors 5.5 GBmodel-00051-of-00097.safetensors 5.5 GBmodel-00041-of-00097.safetensors 5.5 GBmodel-00050-of-00097.safetensors 5.5 GBmodel-00085-of-00097.safetensors 5.5 GBmodel-00055-of-00097.safetensors 5.5 GBmodel-00025-of-00097.safetensors 5.5 GBmodel-00017-of-00097.safetensors 5.5 GBmodel-00086-of-00097.safetensors 5.5 GBmodel-00074-of-00097.safetensors 5.5 GBmodel-00096-of-00097.safetensors 5.5 GBmodel-00034-of-00097.safetensors 5.5 GBmodel-00062-of-00097.safetensors 5.5 GBmodel-00007-of-00097.safetensors 5.5 GBmodel-00081-of-00097.safetensors 5.5 GBmodel-00061-of-00097.safetensors 5.5 GBmodel-00022-of-00097.safetensors 5.5 GBmodel-00060-of-00097.safetensors 5.5 GBmodel-00035-of-00097.safetensors 5.5 GBmodel-00066-of-00097.safetensors 5.5 GBmodel-00065-of-00097.safetensors 5.5 GBmodel-00031-of-00097.safetensors 5.5 GBmodel-00077-of-00097.safetensors 5.5 GBmodel-00054-of-00097.safetensors 5.5 GBmodel-00057-of-00097.safetensors 5.5 GBmodel-00056-of-00097.safetensors 5.5 GBmodel-00072-of-00097.safetensors 5.5 GBmodel-00043-of-00097.safetensors 5.5 GBmodel-00003-of-00097.safetensors 5.5 GBmodel-00058-of-00097.safetensors 5.5 GBmodel-00046-of-00097.safetensors 5.5 GBmodel-00012-of-00097.safetensors 5.5 GBmodel-00030-of-00097.safetensors 5.5 GBmodel-00083-of-00097.safetensors 5.4 GBmodel-00070-of-00097.safetensors 5.4 GBmodel-00006-of-00097.safetensors 5.4 GBmodel-00021-of-00097.safetensors 5.4 GBmodel-00067-of-00097.safetensors 5.4 GBmodel-00095-of-00097.safetensors 5.4 GBmodel-00008-of-00097.safetensors 5.4 GBmodel-00076-of-00097.safetensors 5.4 GBmodel-00011-of-00097.safetensors 5.4 GBmodel-00015-of-00097.safetensors 5.4 GBmodel-00091-of-00097.safetensors 5.4 GBmodel-00027-of-00097.safetensors 5.4 GBmodel-00090-of-00097.safetensors 5.4 GBmodel-00059-of-00097.safetensors 5.4 GBmodel-00036-of-00097.safetensors 5.4 GBmodel-00053-of-00097.safetensors 5.4 GBmodel-00013-of-00097.safetensors 5.4 GBmodel-00001-of-00097.safetensors 5.3 GBmodel-00097-of-00097.safetensors 334 MBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo pipenetwork/Inkling-MLX-4bit ./model-folderthinkingmachines/Inkling