TiGa-RCE/needle-mlx

Needle MLX

An MLX safetensors format conversion of Cactus-Compute/needle, a 26M-parameter encoder-decoder function-calling model.

This is not a fine-tune or a quantization. The published checkpoint is loaded using Needle's declared bfloat16 inference dtype, then stored as MLX safetensors so it preserves the upstream runtime's executable values.

What is included

• model.safetensors: 31 converted tensors, 26,315,421 parameters.
• needle_mlx.py: a small custom MLX inference implementation for Needle's architecture.
• tokenizer/: the upstream SentencePiece tokenizer files.
• config.json and manifest.json: architecture and conversion provenance.

Use

from needle_mlx import NeedleModel

model = NeedleModel.from_pretrained(".")
logits = model.forward([[1, 2, 3]], [[1, 4]])

This is an MLX-native package, not an MLX-LM or oMLX model yet. Needle uses a custom JAX/Flax architecture, so load it with the included runtime or add a dedicated adapter to the serving runtime.

Conversion and verification

The published needle.pkl checkpoint was read with a restricted NumPy-only pickle loader, converted to MLX bfloat16 safetensors, and checked tensor-for-tensor after the same dtype cast used by Needle's upstream runtime.

• Source SHA-256: 40a32e91d1d4197bf15ba559b74f6727c342dc8746918742fc7d8e2c1f18df40
• Converted model.safetensors SHA-256: 7b9d5f0d6ddeb7fbb20f4e45f3f616919357e5d08b5778859fdb762a33d60dae
• Verification: 31/31 tensors equal the source after Needle's bfloat16 runtime cast; an MLX encoder-decoder smoke pass produced logits with shape (1, 2, 8192) and generated the upstream weather tool-call example.

The converter source is available at seeker-cyber-maker/needle-mlx-depicklinator.

Credits and license

Needle was created by Cactus Compute. See the upstream model card and source repository for the model, training details, and citation. The upstream model card publishes Needle under the MIT license; this converted package retains that license.