hunch-0.6b-preview-GGUF

The GGUF conversion of antareslabs/hunch-0.6b-preview at the precisions that passed the equivalence gate: every file was scored on all 6,000 held-out questions against the fp32 PyTorch run of the same checkpoint.

hunch-0.6b-preview.f16.gguf: changes 1 of 6,000 answers against fp32 (bf16 alone changes 28), max TV 4.88e-03 against a floor of 7.68e-02; read on Apple silicon (Metal); gate: pass.

The pass rule and the builds that failed it are in FORMATS. Load it with hunch.formats.hunch_gguf.load(path) from the Hunch repository; the release temperature is embedded in the file and applied by default.

Files: hunch-0.6b-preview.f16.gguf.

License: Apache-2.0, as for the model it converts (model card).

Downloads last month: 75

GGUF

Model size

0.6B params

Architecture

qwen3

Hardware compatibility

16-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for antareslabs/hunch-0.6b-preview-GGUF

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Finetuned

antareslabs/hunch-0.6b-preview

Quantized

(1)

this model