Inkling 2 Mini is a multimodal looped mixture-of-experts model from Thinking Machines Lab: 26.4B parameters, 3.4B active (5.4B effective with three passes through the looped middle layers), text and image input, up to 1M tokens of context. This repository is an unmodified quantization of the official checkpoint (revision 1fb4134).
GGUF IQ3_XXS (importance-matrix calibrated) for llama.cpp and LM Studio, with the vision projector. The smallest build here, for GPUs and Macs with 16 GB of memory.
Other formats from INTERPRETIC AI: GGUF Q4_K_M · GGUF IQ3_XXS · MLX 4-bit
| File | Description | Size |
|---|---|---|
Inkling-2-Mini-IQ3_XXS.gguf |
Language model, IQ3_XXS | 11.7 GB |
mmproj-Inkling-2-Mini-F16.gguf |
Vision projector, F16 | 0.64 GB |
SHA-256 checksums are listed in SHA256SUMS.
Requires llama.cpp with Inkling 2 Mini support (commit 29b52d3 or later).
llama-server -m Inkling-2-Mini-IQ3_XXS.gguf \
--mmproj mmproj-Inkling-2-Mini-F16.gguf \
-ngl 999 -c 32768 --jinja
The --mmproj file is only needed for image input. Measured on one RTX PRO 6000: about 86 tokens/s single-stream generation, 15.5 GB of GPU memory at 16K context. IQ3_XXS trades some quality for size; use the Q4_K_M build if it fits.
Apache 2.0, as the original model. See the model card of Thinking Machines Lab and its acceptable use policy.
INTERPRETIC AI · interpretic.ai
thinkingmachines/Inkling-2-Mini