Index Team publishes official GGUF builds for its 35B-A3B translation model

Index Team has published an official GGUF conversion of its Index-Translate-35B-A3B-preview model, adding a first-party local-inference path for the multilingual translation family. The repository describes Index-Translate as covering 150 languages and supporting terminology- and format-constrained translation as well as controlled dubbing and long-document workflows. Hugging Face marks the GGUF repository Apache-2.0, and the model tree links it directly to the non-GGUF preview checkpoint rather than presenting it as a separate base architecture.

The artifact set spans a wide range of static post-training quantizations produced with llama.cpp. Index Team lists Q2_K at 13.25 GB, Q3 variants from 15.55 GB, Q4_K_M at 21.71 GB, Q5 variants around 25 GB, Q6_K at 29.21 GB, Q8_0 at 37.80 GB and a two-shard F16 conversion baseline at 71.07 GB. The repository also contains multimodal projector files for image-input use; text-only translation can use the plain GGUF files. The model card provides direct llama.cpp serving commands and recommends Q4_K_M as a balance point.

Index Team says each quantization tier was checked against the F16 conversion using GPU validation and generation spot checks before the October 3 publication. Those validation statements remain publisher claims. For the registry, the important distinction is artifact lineage: these are official quantized derivatives of the preview model, so exact GGUF files and formats can be tracked without collapsing the quantized repository and its source checkpoint into unrelated model records.