Index-Translate-35B-A3B-preview-GGUF

Official GGUF conversion of IndexTeam/Index-Translate-35B-A3B-preview, part of the Index-Translate multilingual translation model family (150 languages, terminology/format-constrained translation, controlled dubbing translation, long-document translation).

• Technical report: Index-Translate: A Multilingual Translation Model Family
• Code: github.com/bilibili/Index-Translate
• Converted with llama.cpp (master, 2026-10); static post-training quantization, all bit-widths in this single repository.

Provided files

(sorted by size)

File	Size	Notes
mmproj-Q8_0	0.61 GB	multi-modal projector supplement
mmproj-f16	0.90 GB	multi-modal projector supplement
Q2_K	13.25 GB	2-bit, significant quality loss
Q3_K_S	15.55 GB	3-bit, noticeable quality loss
Q3_K_M	17.17 GB	3-bit, noticeable quality loss
Q3_K_L	18.55 GB	3-bit, noticeable quality loss
IQ4_XS	19.39 GB	4-bit I-quant, small
Q4_K_S	20.37 GB	4-bit, slightly smaller/faster
Q4_K_M	21.71 GB	4-bit, good balance, recommended
Q5_K_S	24.56 GB	5-bit, low quality loss
Q5_K_M	25.35 GB	5-bit, low quality loss
Q6_K	29.21 GB	6-bit, very low quality loss
Q8_0	37.80 GB	8-bit, near-lossless
f16	71.07 GB	16-bit, lossless conversion baseline, 2 shards

Multimodal Image Translation (mmproj)

This repository includes official native multi-modal projector weights (*.mmproj-*.gguf), enabling end-to-end image-to-text translation across 150 languages without any external OCR tools:

• Index-Translate-35B-A3B-preview.mmproj-Q8_0.gguf: 8-bit quantized projector, recommended for production and local use.
• Index-Translate-35B-A3B-preview.mmproj-f16.gguf: 16-bit lossless conversion baseline.

1. Launch llama-server with Vision Projector:

llama-server \
  -m Index-Translate-35B-A3B-preview.Q4_K_M.gguf \
  --mmproj Index-Translate-35B-A3B-preview.mmproj-Q8_0.gguf \
  -ngl 99 \
  -c 32768 \
  --port 8000 \
  --alias Index-Translate-35B-A3B-preview

2. Run Image Translation:

Once the server is running, perform image translation with the official script:

python inference/llm/translate.py --image screenshot.png --target zh -m Index-Translate-35B-A3B-preview

Or test it directly via the online web demo at https://index-translate.bilibili.com/?p=/site/translate.html. Full guide: Multimodal Image Translation Guide.

Usage

# Run the recommended Q4_K_M directly:
llama serve -hf IndexTeam/Index-Translate-35B-A3B-preview-GGUF:Q4_K_M
llama cli -hf IndexTeam/Index-Translate-35B-A3B-preview-GGUF:Q4_K_M

Translation prompt format (greedy decoding, temperature=0 recommended; use the chat template with enable_thinking: false):

请将以下文本翻译为{target-language}，直接输出翻译结果，不要进行任何解释。

{source-text}

# llama.cpp: run the recommended Q4_K_M with the chat template and thinking disabled
llama-cli -hf IndexTeam/Index-Translate-35B-A3B-preview-GGUF:Q4_K_M --jinja --chat-template-kwargs '{"enable_thinking":false}' --temp 0

Prompting & constrained translation (instTrans)

Beyond plain translation, the models follow the instTrans constrained-translation format. The official client wraps requests into the canonical structure 【源文】<text> + numbered 1. 【硬性要求】<hard constraints> + 2. 【注意】<soft constraints> + suffix instructions:

• Hard constraints (binary, must hold): strict terminology glossary enforcement (e.g. 碳纤维:carbon fiber, 抗裂缝:crack resistance), and format/structure preservation for JSON/CSV/code/placeholders.
• Soft constraints (graded): tone & style adaptation (e.g. formal business-email register), domain/word-sense disambiguation (e.g. plant → 工厂 in an industrial context), cross-sentence consistency, LaTeX preservation.
• Syllable-controlled translation (dubbing): use the Index-Homura checkpoints (IndexTeam/Index-Homura-2B/9B), which strictly respect a target syllable budget and can be combined with glossaries.

Full prompt reference: github.com/bilibili/Index-Translate (Instruction Following section, docs/prompts.md, inference/llm/cases/).

See the base model card for the full instTrans constrained-translation format and serving presets.

Consistency validation

Before release, each quantization tier was validated on GPU (NVIDIA A100) against the F16 conversion: per-token KL divergence / RMS Δp via llama-perplexity, plus greedy-generation spot checks against the original BF16 weights (transformers reference). Generation outputs of Q4_K_M matched the reference almost verbatim.

Converted and published by the Index team, 2026-10-03.