A Chinese Video Platform Just Out-Translated Google's Gemini With a 3B-Parameter Model
On September 30, Bilibili's Index LLM team open-sourced Index-Translate, a family of translation models built on Alibaba's Qwen3.5 base and released under an Apache 2.0 license on Hugging Face and ModelScope. On its face, it's a niche release from a company known for anime fandom and livestreamed gaming, not a frontier AI lab. But the benchmark numbers the team published, and the product problem the models are built to solve, make Index-Translate worth more scrutiny than its low profile has gotten so far.
Built for a video platform's actual problem, not a leaderboard
Most open translation efforts, from Meta's 2022 NLLB-200 to the translation modes bolted onto general chat models, optimize for raw sentence-pair accuracy across as many languages as possible. Index-Translate instead ships as four purpose-built variants: a general text model, a speech-to-speech dubbing model (Index-Echo) that preserves a speaker's vocal characteristics, a syllable-constrained dubbing model (Index-Homura) that fits translated dialogue into the original line's timing, and a long-document model (Index-NativeLong) that handles full novels in a single pass with context windows up to 262,144 tokens. That's a product roadmap shaped by Bilibili's own business: a user-generated video platform that needs subtitles and dubbing at scale across 150 languages, not a benchmark-chasing research exercise.
Small active-parameter counts, frontier-level scores
The headline number is the 35B-A3B variant, a mixture-of-experts model with 35 billion total parameters but only about 3 billion active per token. On the FLORES translation benchmark it scores 0.8794, edging out Google's Gemini 3.5 Flash Lite at 0.8750, and on instruction-compliance scoring it roughly matches DeepSeek's V4.1-Flash. The syllable-constrained Homura-9B variant hits 81.92% accuracy at keeping dubbed lines within 10% of the original line's syllable count on Bilibili's internal SandGlass benchmark, a dubbing-specific metric that general-purpose chat models simply aren't trained to hit. None of these are frontier-scale wins in isolation, but the pattern is the story: a model running a fraction of Gemini's active compute is matching or beating it on the specific task it was purpose-built for.
"Real content demands more from translation than simply 'switching languages,'" Bilibili's Index LLM team writes on the project's site, arguing that a joke should keep its meaning, a dubbed line should fit the original rhythm, and a novel needs a coherent, complete translation rather than a literal one.
Why this matters beyond one company's subtitle problem
Index-Translate lands in the same week as a cluster of frontier releases: Anthropic's Claude Sonnet 5.5, OpenAI's cheaper GPT-6.1 Sol, and Google's cautiously gated Gemini 4 Argon, all competing on general-purpose reasoning and coding benchmarks at enormous compute cost. Bilibili's release is a quiet counter-signal to that race. It's evidence that companies sitting on a specific, high-volume use case, in this case a video platform needing localization at scale, can fine-tune a mid-sized open base model into something that beats a trillion-dollar lab's general model on the one task that matters to them, at a fraction of the inference cost. That's a different kind of AI competition than the one dominating headlines: not who builds the biggest or smartest model, but who builds the most useful one for a specific job. It also extends a pattern already visible in China's open-model ecosystem, where Qwen derivatives increasingly serve as the base layer that product companies, not just AI labs, build specialized systems on top of.
What to watch next: whether Index-Translate's 35B-A3B variant graduates from preview to a full release with independently verified benchmarks, since the FLORES and instruction-compliance numbers above are self-reported by Bilibili's team; whether other consumer platforms with heavy localization needs, YouTube, TikTok, Spotify, follow with their own task-specific fine-tunes rather than licensing general-purpose APIs; and whether this kind of specialist release starts showing up in Hugging Face's trending charts alongside the quantized frontier-model derivatives that have dominated open-source activity through the fall.