SenseNova-SI-InternVL3-8B-AWQ

This model was converted to AWQ format from sensenova/SenseNova-SI-InternVL3-8B using lmdeploy via the following way:

conda create -n lmdeploy python=3.10 -y
conda activate lmdeploy
pip install lmdeploy "transformers>=4.53.0,<4.54" timm datasets
lmdeploy lite auto_awq sensenova/SenseNova-SI-InternVL3-8B --work-dir SenseNova-SI-InternVL3-8B-AWQ --dtype bfloat16

Then modify the config.json, put quantization_config at the outermost level.

{
  "architectures": [
    "InternVLChatModel"
  ],
  "auto_map": {
    "AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
    "AutoModel": "modeling_internvl_chat.InternVLChatModel",
    "AutoModelForCausalLM": "sensenova/SenseNova-SI-InternVL3-8B--modeling_internvl_chat.InternVLChatModel"
  },
  "downsample_ratio": 0.5,
  "dynamic_image_size": true,
  "force_image_size": 448,
  "hidden_size": 3584,
  "image_fold": null,
  "llm_config": {
    "_name_or_path": "./pretrained/Qwen2.5-32B-Instruct",
    "architectures": [
      "Qwen2ForCausalLM"
    ],
    "attention_dropout": 0.0,
    "attn_implementation": "eager",
    "bos_token_id": 151643,
    "eos_token_id": 151643,
    "hidden_act": "silu",
    "hidden_size": 3584,
    "initializer_range": 0.02,
    "intermediate_size": 18944,
    "layer_types": [
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention",
      "full_attention"
    ],
    "max_position_embeddings": 32768,
    "max_window_layers": 70,
    "model_type": "qwen2",
    "moe_config": null,
    "num_attention_heads": 28,
    "num_hidden_layers": 28,
    "num_key_value_heads": 4,
    "rms_norm_eps": 1e-06,
    "rope_scaling": {
      "factor": 2.0,
      "rope_type": "dynamic",
      "type": "dynamic"
    },
    "rope_theta": 1000000.0,
    "sliding_window": null,
    "torch_dtype": "bfloat16",
    "use_bfloat16": true,
    "use_cache": false,
    "use_sliding_window": false,
    "vocab_size": 151674
  },
  "quantization_config": {
    "bits": 4,
    "group_size": 128,
    "quant_method": "awq",
    "version": "gemm",
    "zero_point": true,
    "modules_to_not_convert": ["vision_model", "mlp1"]
  },
  "max_dynamic_patch": 12,
  "min_dynamic_patch": 1,
  "model_type": "internvl_chat",
  "output_attentions": false,
  "pad2square": false,
  "ps_version": "v2",
  "select_layer": -1,
  "system_message": null,
  "template": "internvl2_5",
  "tie_word_embeddings": false,
  "torch_dtype": "float16",
  "transformers_version": null,
  "use_backbone_lora": 0,
  "use_llm_lora": 0,
  "use_thumbnail": true,
  "vision_config": {
    "_name_or_path": "OpenGVLab/InternViT-6B-448px-V1-5",
    "architectures": [
      "InternVisionModel"
    ],
    "attention_dropout": 0.0,
    "auto_map": {
      "AutoConfig": "configuration_intern_vit.InternVisionConfig",
      "AutoModel": "modeling_intern_vit.InternVisionModel"
    },
    "capacity_factor": 1.2,
    "drop_path_rate": 0.0,
    "dropout": 0.0,
    "eval_capacity_factor": 1.4,
    "hidden_act": "gelu",
    "hidden_size": 1024,
    "image_size": 448,
    "initializer_factor": 0.1,
    "initializer_range": 1e-10,
    "intermediate_size": 4096,
    "laux_allreduce": "all_nodes",
    "layer_norm_eps": 1e-06,
    "model_type": "intern_vit_6b",
    "moe_coeff_ratio": 0.5,
    "moe_intermediate_size": 768,
    "moe_output_scale": 4.0,
    "noisy_gate_policy": "RSample_before",
    "norm_type": "layer_norm",
    "num_attention_heads": 16,
    "num_channels": 3,
    "num_experts": 8,
    "num_hidden_layers": 24,
    "num_routed_experts": 4,
    "num_shared_experts": 4,
    "patch_size": 14,
    "qk_normalization": false,
    "qkv_bias": true,
    "shared_expert_intermediate_size": 3072,
    "torch_dtype": "bfloat16",
    "use_bfloat16": true,
    "use_flash_attn": false,
    "use_moe": false,
    "use_residual": true,
    "use_rts": false,
    "use_weighted_residual": false
  }
}

SenseNova-SI: Scaling Spatial Intelligence with Multimodal Foundation Models

[EASI Codebase] [EASI Leaderboard]

Overview

Despite remarkable progress, leading multimodal models still exhibit notable deficiencies in spatial intelligence: the ability to make metric estimations, understand spatial relationships, handle viewpoint changes, and integrate information across complex scenes. We take a scaling perspective: constructing and curating a large-scale, comprehensive collection of spatial intelligence data, and through continued training on powerful multimodal foundations, cultivating multi-faceted spatial understanding within the SenseNova-SI family of models. In the future, SenseNova-SI will be integrated with larger-scale in-house models.

Release Information

Currently, we build SenseNova-SI upon popular open-source foundation models to maximize compatibility with existing research pipelines. In this release, we present SenseNova-SI-InternVL3-2B and SenseNova-SI-InternVL3-8B, which achieve state-of-the-art performance among open-source models of comparable size across four recent spatial intelligence benchmarks: VSI, MMSI, MindCube, and ViewSpatial.

<table> <thead> <tr> <th>Model</th> <th>VSI</th> <th>MMSI</th> <th>MindCube-Tiny</th> <th>ViewSpatial</th> </tr> </thead> <tbody> <tr style="background:#F2F0EF;font-weight:700;text-align:center;"> <td colspan="5"><em>Open-source Models (~2B)</em></td> </tr> <tr> <td>InternVL3-2B</td><td>32.98</td><td>26.50</td><td>37.50</td><td>32.56</td> </tr> <tr> <td>Qwen3-VL-2B-Instruct</td><td>50.36</td><td>28.90</td><td>34.52</td><td>36.97</td> </tr> <tr> <td>MindCube-3B-RawQA-SFT</td><td>17.24</td><td>1.70</td><td>51.73</td><td>24.14</td> </tr> <tr> <td>MindCube-3B-Aug-CGMap-FFR-Out-SFT</td><td>29.60</td><td>29.10</td><td>41.06</td><td>30.90</td> </tr> <tr> <td>MindCube-3B-Plain-CGMap-FFR-Out-SFT</td><td>29.93</td><td>30.40</td><td>39.90</td><td>31.20</td> </tr> <tr> <td>SpatialLadder-3B</td><td>44.86</td><td>27.40</td><td>43.46</td><td>39.85</td> </tr> <tr> <td>SpatialMLLM-4B</td><td>45.98</td><td>26.10</td><td>33.46</td><td>34.66</td> </tr> <tr> <td><strong>SenseNova-SI-InternVL3-2B</strong></td> <td><strong>58.47</strong></td> <td><strong>35.50</strong></td> <td><strong>71.35</strong></td> <td><strong>40.62</strong></td> </tr> <tr style="background:#F2F0EF;font-weight:700;text-align:center;"> <td colspan="5"><em>Open-source Models (~8B)</em></td> </tr> <tr> <td>InternVL3-8B</td><td>42.14</td><td>28.00</td><td>41.54</td><td>38.66</td> </tr> <tr> <td>Qwen3-VL-8B-Instruct</td><td>57.90</td><td>31.10</td><td>29.42</td><td>42.20</td> </tr> <tr> <td>BAGEL-7B</td><td>30.90</td><td>33.10</td><td>34.71</td><td>41.32</td> </tr> <tr> <td>SpaceR-7B</td><td>36.29</td><td>27.40</td><td>37.98</td><td>35.85</td> </tr> <tr> <td>ViLaSR-7B</td><td>44.63</td><td>30.20</td><td>35.10</td><td>35.71</td> </tr> <tr> <td><strong>SenseNova-SI-InternVL3-8B</strong></td> <td><strong>62.80</strong></td> <td><strong>37.90</strong></td> <td><strong>89.33</strong></td> <td><strong>53.92</strong></td> </tr> <tr style="background:#F2F0EF;color:#6b7280;font-weight:600;text-align:center;"> <td colspan="5"><em>Proprietary Models</em></td> </tr> <tr style="color:#6b7280;"> <td>Gemini-2.5-pro-2025-06</td><td>53.57</td><td>38.00</td><td>57.60</td><td>46.06</td> </tr> <tr style="color:#6b7280;"> <td>Grok-4-2025-07-09</td><td>47.92</td><td>37.80</td><td>63.56</td><td>43.23</td> </tr> <tr style="color:#6b7280;"> <td>GPT-5-2025-08-07</td><td>55.03</td><td>41.80</td><td>56.30</td><td>45.59</td> </tr> </tbody> </table>