inclusionAILanguage modelActive
ZwZ2-9B
ZwZ2-9B has no overall index value yet because there are too few independent benchmark results.
Summary
ZwZ2-9B is the Qwen3.5-9B variant of Zooming without Zooming v2. It is designed to connect visual evidence across image regions. inclusionAI trained the family with GRPO on about 20,400 verified multi-region questions built from SA-1B images. Privileged crop views were used to construct the data; the trained model receives the full image and needs no zooming tools.
The sources report 25 benchmark results across six domains, six domain means, and the overall mean. The model card reports gains over the base model on all 25 benchmarks. The paper also publishes eight top-k evidence IoU measurements on TreeBench examples for 9B. The 7K ablations are not attributed to this released checkpoint.
The paper evaluates models in either low-thinking or non-thinking mode but does not assign a precise mode to each benchmark. Hugging Face model metadata does not state a weight license; the code repository's Apache-2.0 license does not automatically apply to the weights.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 40 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.
VStar (Show measurement, test conditions, and source)
- Source value
- 90.58
- Score
- 90.58
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
HR-4K (Show measurement, test conditions, and source)
- Source value
- 89.75
- Score
- 89.75
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
HR-8K (Show measurement, test conditions, and source)
- Source value
- 87.62
- Score
- 87.62
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
MME-RealWorld EN (Show measurement, test conditions, and source)
- Source value
- 75.33
- Score
- 75.33
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
MME-RealWorld CN (Show measurement, test conditions, and source)
- Source value
- 73.45
- Score
- 73.45
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
ZoomBench (Show measurement, test conditions, and source)
- Source value
- 59.53
- Score
- 59.53
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
Chartography (Show measurement, test conditions, and source)
- Source value
- 13
- Score
- 13
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
PerceptionBench (Show measurement, test conditions, and source)
- Source value
- 38.89
- Score
- 38.89
- Metric
- Accuracy
- Unit
- %
- Category
- Fine-grained perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
TreeBench (Show measurement, test conditions, and source)
- Source value
- 51.6
- Score
- 51.6
- Metric
- Accuracy
- Unit
- %
- Category
- Spatial perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
O3-Bench (Show measurement, test conditions, and source)
- Source value
- 67.83
- Score
- 67.83
- Metric
- Accuracy
- Unit
- %
- Category
- Spatial perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
CountQA (Show measurement, test conditions, and source)
- Source value
- 48.43
- Score
- 48.43
- Metric
- Accuracy
- Unit
- %
- Category
- Spatial perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
MIMIC (Show measurement, test conditions, and source)
- Source value
- 64.98
- Score
- 64.98
- Metric
- Accuracy
- Unit
- %
- Category
- Spatial perception
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- ZwZ2 paper (arXiv:2610.37037), Table 1
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Accuracy (%), Table 1. The paper says models were evaluated in low-thinking or non-thinking modes but does not map a precise mode to each benchmark. No per-benchmark shot count is reported. The three external API reference models had platform-limited input resolution; exact limits are not listed.
- Context
- Author-reported accuracy. The paper does not give a per-benchmark sample count or confidence interval.
Page 1 of 4
All specifications
Architecture
- Architecture
- Qwen3.5/Qwen3.6 multimodal model. Configuration: 32 language layers, hidden size 4096, 16 attention heads, intermediate size 12288; vision encoder with 27 layers and hidden size 1152.
- Source: Architecture
Input and output
Further details
- Configuration position limit
- The nested text configuration sets max_position_embeddings to 262,144. This is a configuration value, not a separate guarantee of the usable length in every runtime.
- Source: Configuration position limit
- Training method
- ZwZ2 was trained with GRPO, without additional supervised fine-tuning, on about 20,400 verified multi-region VQA examples. The examples were built from SA-1B images; training uses the full image rather than the privileged crop views.
- Source: Training method
- Reasoning mode
- Training rollouts used non-thinking mode. The published inference example sets enable_thinking=false and do_sample=false.
- Source: Reasoning mode
- Benchmark coverage
- The model card publishes the complete results table with 25 benchmarks across six domains plus domain and overall means.
- Source: Benchmark coverage
- Weight license
- The Hugging Face model metadata does not state a license. The card directs users to the applicable weight and backbone licenses.
- Source: Weight license
- Code repository license
- The separate ZwZ2 code repository is licensed under Apache-2.0. That code license does not establish a license for the model weights.
- Source: Code repository license
- Parameter count reporting
- The model name and base model identify the 9B scale; the HF page sidebar also lists 9B parameters. config.json does not give a more precise total.
- Source: Parameter count reporting
- Code repository license
- Apache-2.0 applies to the code repository, not automatically to the model weights.
- Source: Code repository license
Frequently asked questions about ZwZ2-9B
How large is the context window of ZwZ2-9B?
262,144 tokens. That is how much text ZwZ2-9B reads at once, about 390 A4 pages.
Page count as a rule of thumb from Artificial Analysis
Can I run ZwZ2-9B locally?
Yes, the weights are open.
Stated by inclusionAI
- Open weights
- yes
- License
- n/a
- Parameters
- 9B
- active per token
- n/a
Sources and data date
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 29, 2026. 61 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | Per-benchmark sample counts and confidence intervals for Table 1 Note The paper reports scores but does not publish per-benchmark n or confidence intervals for Table 1 evaluations. |
| Not published | Exact benchmark mode and resolution by benchmark Note The paper says low-thinking or non-thinking and notes platform-limited resolution for three external API baselines. It does not give a complete per-model, per-benchmark mapping or exact resolution values. |
| Not published | Exact point values in Figure 5 context-length curves Note Figure 5 plots MemLens and MMLongBench visual-core curves without tabulated point labels. No values are estimated from the graph. |
| Not published | TreeBench layer-wise attention IoU for 27B Note Table 5 compares only 4B and 9B against their base models; 27B layer-wise attention values are not published. |
| Not found in the inspected sources | Exact public weight release date Note The paper and model cards were public by the verification date, but an exact first-public date for the weight files was not verified. HF repository creation timestamps are not treated as release dates. |
| Not found in the inspected sources | Weight license Note HF model metadata declares no license. The model card points to separate weight and base-model terms; Apache-2.0 for the code does not establish the model-weight license. |
| Not found in the inspected sources | Vals rows in captured payloads Note Nine complete Vals payloads dated 28 October were checked; no exact model ID or unambiguous model label. Limited to these saved benchmarks. |
| Unresolved | Uncaptured Arena Code/Speech leaderboards Note Text, Agent including Code/Chat/Work, and twelve WebDev views were captured. A separate general Code and Arena Speech leaderboard were not saved as complete tables. |
| Unresolved | Artificial Analysis catalog reconciliation Note The full current AA model and benchmark catalog was not re-reconciled for this profile in this sidecar. This sidecar makes no AA absence claim; native AA imports remain authoritative. |
| Additional source | ZwZ2-9B official model card (retrieved October 29, 2026) · ZwZ2-9B official model card · Editorial description reviewed October 29, 2026 |
| Additional source | ZwZ2-9B published configuration (retrieved October 29, 2026) · ZwZ2-9B configuration · Editorial description reviewed October 29, 2026 |
| Additional source | ZwZ2 full-resolution results image (retrieved October 29, 2026) · ZwZ2-9B full-resolution results image · Editorial description reviewed October 29, 2026 |
| Additional source | ZwZ2 paper, arXiv:2610.37037 (retrieved October 29, 2026) |
| Additional source | ZwZ2 project repository (retrieved October 29, 2026) · ZwZ2 project repository · Editorial description reviewed October 29, 2026 |
| Additional source | Model metadata (retrieved October 29, 2026)12 |
| Additional source | ZwZ2 paper · Editorial description reviewed October 29, 2026 |