Bilibili Index-Translate

Index-Translate, released September 30, 2026 by Bilibili's Index LLM Team under Apache 2.0, is a family of translation-first models covering 150 languages. The flagship is Index-Translate-35B-A3B (preview): 35 billion total parameters with only 3 billion active per forward pass, so inference costs sit near a dense 3B model. Dense 2B and 9B variants cover simpler deployments.

The family follows translation instructions (terminology, formatting, content preservation) and extends to speech: Index-Echo translates subtitles and dubbed speech while preserving the speaker's voice characteristics, and Index-Homura steers translations toward a requested syllable count for dubbing. Index-NativeLong translates whole documents with cross-passage context. Weights ship in GGUF, FP8, and NVFP4 quantized builds, and Bilibili runs a free OpenAI-compatible public API.

On the team's benchmarks, the 35B-A3B preview scores 0.8794 on FLORES (COMET-22) and 76.76 on a WMT26 judge score; the 9B model leads the team's low-resource instruction test with the lowest off-target rate at 3.47%.

Provider-reported (Index Team): 35B-A3B preview 0.8794 FLORES COMET-22, 76.76 WMT26 judge; 9B best low-resource instruction score 0.7725 with 3.47% off-target rate; competitive with DeepSeek-V4.1-Flash and GPT-5.6-Sol on translation axes. Independent: no published independent benchmark scores at research time.