AppleSpecialist modelOpen weights
LensVLM-9B
20out of 100new calculation
based on two data groups, mostly search, range 10 to 29
Summary
LensVLM-9B is Apple’s vision-language model for long documents. It scans compressed visual representations of text and expands only the pages relevant to the current question. Visual processing therefore stays selective instead of loading every page at full resolution.
The 9-billion-parameter checkpoint runs locally with Transformers, vLLM, or SGLang under Apple’s Machine Learning Research Model License. The released demo supports 5x, 10x, and 15x compression. Its model card lists no hosted Inference Provider.
On this page
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 42 published benchmark results
Page 1 of 4
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.
LensVLM document QA accuracy, 5x (Show measurement, test conditions, and source)
- Source value
- 68.9
- Metric
- percent
- Unit
- %
- Category
- document-qa
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 5x input compression; LLM-judge semantic accuracy; unweighted macro-average over seven datasets.
LensVLM document QA accuracy, 10x (Show measurement, test conditions, and source)
- Source value
- 62.1
- Metric
- percent
- Unit
- %
- Category
- document-qa
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 10x input compression; LLM-judge semantic accuracy; unweighted macro-average over seven datasets.
LensVLM document QA accuracy, 15x (Show measurement, test conditions, and source)
- Source value
- 52.1
- Metric
- percent
- Unit
- %
- Category
- document-qa
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 15x input compression; LLM-judge semantic accuracy; unweighted macro-average over seven datasets.
LensVLM evidence-page selection, 5x (Show measurement, test conditions, and source)
- Source value
- 76.8
- Metric
- percent
- Unit
- %
- Category
- document-retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 5x input compression; ground-truth evidence-page hit.
LensVLM evidence-page selection, 10x (Show measurement, test conditions, and source)
- Source value
- 71
- Metric
- percent
- Unit
- %
- Category
- document-retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 10x input compression; ground-truth evidence-page hit.
LensVLM evidence-page selection, 15x (Show measurement, test conditions, and source)
- Source value
- 52.1
- Metric
- percent
- Unit
- %
- Category
- document-retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LensVLM technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Final LensVLM SFT+RL; 15x input compression; ground-truth evidence-page hit.
LensVLM QA accuracy on NQ, 5x (Show measurement, test conditions, and source)
- Source value
- 76.1
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 5x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
LensVLM QA accuracy on NQ, 10x (Show measurement, test conditions, and source)
- Source value
- 75.1
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 10x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
LensVLM QA accuracy on NQ, 15x (Show measurement, test conditions, and source)
- Source value
- 71.2
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 15x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
LensVLM QA accuracy on HotpotQA, 5x (Show measurement, test conditions, and source)
- Source value
- 81.5
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 5x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
LensVLM QA accuracy on HotpotQA, 10x (Show measurement, test conditions, and source)
- Source value
- 75.5
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 10x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
LensVLM QA accuracy on HotpotQA, 15x (Show measurement, test conditions, and source)
- Source value
- 64.5
- Metric
- LLM-judged semantic QA accuracy
- Unit
- %
- Direction
- Higher is better
- Sample
- 600
- Source type
- vendor-reported
- Evaluator
- Apple, LensVLM technical paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. LensVLM SFT+RL on Qwen3.5-9B-Base; Qwen3.5-397B judge; 15x input compression; T=6; mean of three runs.
- Context
- The paper reports a three-run mean and no confidence interval.
Page 1 of 4
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 21 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 20, 2026
LensVLM-9BBielik 11B v3.1 DFlash2 22Hy-SimulMT-30B-A3B 22ManiFlow-tb 21Molmo2-4B-tb 21
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
- 2November 1, 2026: Method version 4 (common scale, shrinkage)
- 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as LensVLM-9B, chosen by model line, developer, size, price and release date.
- AstaBrief-8BAllen Institute for AI · similar size
- Yuanbao-Reranker-8BTencent · similar size
- pplx-embed-v2-context-9bPerplexity · similar size
- pplx-embed-v2-late-9bPerplexity · similar size
- FLUX 3 ActionBlack Forest Labs · similar size
- Molmo2-4B-tbAi2 · similar size
- Qwen-Drive-1.0-4BQwen · similar size
- Yuanbao-Reranker-4BTencent · similar size
All specifications
Architecture
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- Shutdown
- None announced
- no source
API
Further details
Sources and data date
Show all 9 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 4, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Additional source | Apple LensVLM model card (retrieved October 3, 2026) · Apple LensVLM model card · Editorial description reviewed October 3, 2026 |
| Additional source | LensVLM research paper (retrieved October 3, 2026) · LensVLM research paper · Editorial description reviewed October 3, 2026 |
| Additional source | LensVLM technical report (retrieved October 4, 2026; October 17, 2026) |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |