this-that-model-1.1-gguf

this-that-model-1.1 GGUF

GGUF conversion of flock-io/this-that-model-1.1, a 1.88B typed decision model. It is used by Unbound's on-device Snake demo, which runs in the browser via wllama and on iOS/Android via llama.rn.

Conversion

python convert_hf_to_gguf.py this-that-model-1.1 --no-mtp --outtype f16
llama-quantize this-that-model-1.1-f16.gguf this-that-model-1.1-Q8_0.gguf Q8_0

--no-mtp is required. The config declares one MTP layer whose weights are not shipped. Without the flag, llama.cpp fails to load the file with missing tensor 'blk.24.attn_norm.weight'.

Usage

This is not a chat model. Build the prompt in the thisthat state-first layout:

Context:
<state>

Question: <question>
Options:
(A) ...
(B) ...
Answer: (

Do not add a BOS token. Run one forward pass, take the next-token probabilities of the option letters A, B, …, and renormalise over those letters only.

Checks

• The f16 file matches the PyTorch reference probabilities to three decimal places.
• On 200 snake items from the model's spatial benchmark (limberc/this-that-spatial-bench), Q8_0 scores 99% on move safety and 94% on food direction. Q4_K_M drops to 87% on food direction, which is why only Q8_0 is published here.

Licence: MIT, same as the original model.