Language modelOpen weights
Index-NativeLong-9B
Bilibili Index Team
- Released
- -
- Data date
- October 7, 2026
- Input
- Complete Chinese, English, or Japanese documents
- Output
- Translated documents in English, Japanese, or Chinese
Index-NativeLong-9B offers the document translation of the 2B variant in a larger checkpoint. It translates the whole document in one step to account for terms and references across distant passages. Its four templates connect Chinese in both directions with English or Japanese.
Specifications and access
| Specification | Value and source |
|---|---|
| Model type | Open weightsSource |
| Model ID | IndexTeam/Index-Nailong-9BSource |
| Model task | Document translation that processes the full textSource |
| Access | Public Hugging Face weights for local executionSource |
| Input modalities | Complete Chinese, English, or Japanese documentsSource |
| Output modalities | Translated documents in English, Japanese, or ChineseSource |
| License | Apache 2.0 according to the Hugging Face model cardSource |
| Translation directions | Chinese to and from English, and Chinese to and from Japanese, using fixed templatesSource |
| Document mode | Translates a complete document in one generation and accounts for distant passagesSource |
| Parameters | 9 billionSource |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
NativeLongBench / GuoFeng (zh→en) / five-length mean (Show measurement, test conditions, and source)
- Source value
- 0.7891
- Score
- 0.7891
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / GuoFeng (zh→en) / 4K (Show measurement, test conditions, and source)
- Source value
- 0.7836
- Score
- 0.7836
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 4K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / GuoFeng (zh→en) / 8K (Show measurement, test conditions, and source)
- Source value
- 0.8074
- Score
- 0.8074
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 8K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / GuoFeng (zh→en) / 16K (Show measurement, test conditions, and source)
- Source value
- 0.808
- Score
- 0.808
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 16K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / GuoFeng (zh→en) / 32K (Show measurement, test conditions, and source)
- Source value
- 0.8084
- Score
- 0.8084
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 32K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / GuoFeng (zh→en) / 64K (Show measurement, test conditions, and source)
- Source value
- 0.7382
- Score
- 0.7382
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 64K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / five-length mean (Show measurement, test conditions, and source)
- Source value
- 0.7683
- Score
- 0.7683
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / 4K (Show measurement, test conditions, and source)
- Source value
- 0.7733
- Score
- 0.7733
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 4K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / 8K (Show measurement, test conditions, and source)
- Source value
- 0.7695
- Score
- 0.7695
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 8K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / 16K (Show measurement, test conditions, and source)
- Source value
- 0.779
- Score
- 0.779
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 16K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / 32K (Show measurement, test conditions, and source)
- Source value
- 0.7756
- Score
- 0.7756
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 32K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
NativeLongBench / BWB Track A3 (zh→en) / 64K (Show measurement, test conditions, and source)
- Source value
- 0.744
- Score
- 0.744
- Metric
- document-SEGALE/COMET (max_size=8)
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact published checkpoint IndexTeam/Index-Nailong-9B (public model name Index-NativeLong-9B); NativeLongBench; translate the complete test passage in one generation with greedy decoding and a 256K-token context. Score is document-SEGALE/COMET (max_size=8), 0–1, higher is better. Corpus direction zh→en; five Chinese-side length groups are equally weighted for the corpus mean. Length subgroup: 64K Chinese-side tokens. For GuoFeng/BWB this counts source Chinese; for General Books it counts the Chinese reference translation. The score covers complete-document generation under the same native protocol; exact subgroup sample count is not published.
- Context
- The report states that these fixed evaluation sets were also used during model development. It reports 274 document windows across the three corpora and five length groups, but does not publish the per-corpus/per-bin sample count; no sampleCount is assigned to these rows.
Page 1 of 2
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-corpus/per-length evaluation counts Original wording The report publishes 274 document windows for the complete three-corpus, five-length evaluation, but not the sample count for each corpus or length-bin score. |
| Not found in the inspected sources | independent exact-checkpoint reproduction Original wording The reviewed official card, evaluation repository, and technical report provide vendor-reported results; they do not cite an independent reproduction of this exact Nailong checkpoint. |
| Additional source | Bilibili Index-Translate repository (retrieved October 7, 2026) · Bilibili Index-Translate repository · Editorial description reviewed October 7, 2026 |
| Additional source | Bilibili IndexTeam technical report, Index-Translate (arXiv v1) (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |