Multimodale Modelle
Immer mehr offene Modelle nehmen neben Text auch Bilder auf. Diese Seite listet jedes multimodale Modell des Korpus mit dem Config-Beleg seiner Einstufung.
Kriterium (faktisch, mechanisch): Die mitgelieferte config.json enthält visionsbezogene Felder — Bild-/Vision-Token-IDs, Vision-Encoder-Einstellungen oder Bild-Patch-Felder. Nichts wird darüber hinaus gefolgert.
| Modell | Labor | Veröffentlicht | Parameter | MoE | Attention | Config-Beleg |
|---|---|---|---|---|---|---|
| Step-3 | StepFun | 2025-07 | 321 Mrd. | Sparsames MoE | MHA | im_end_token im_patch_token im_start_token +3 |
| Hunyuan 7B | Hunyuan | 2025-08 | 7.8 Mrd. | Dicht | GQA | im_end_id im_newline_id im_start_id |
| Kimi K2.5 | Kimi | 2026-02 | 1,040.2 Mrd. | Sparsames MoE | MLA | use_unified_vision_chunk |
| Qwen3.5 122B A10B | Qwen | 2026-02 | 119.9 Mrd. | Sparsames MoE | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.5 27B | Qwen | 2026-02 | 23.7 Mrd. | Dicht | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.5 35B A3B | Qwen | 2026-02 | 33.9 Mrd. | Sparsames MoE | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.5 397B A17B | Qwen | 2026-02 | 392.6 Mrd. | Sparsames MoE | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.5 9B | Qwen | 2026-02 | 7.9 Mrd. | Dicht | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.5 9B Base | Qwen | 2026-02 | 7.9 Mrd. | Dicht | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Kimi K2.6 | Kimi | 2026-04 | 1,040.2 Mrd. | Sparsames MoE | MLA | use_unified_vision_chunk |
| Qwen3.6 27B | Qwen | 2026-04 | 23.7 Mrd. | Dicht | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.6 35B A3B | Qwen | 2026-04 | 33.9 Mrd. | Sparsames MoE | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Hunyuan-TurboS 7B | Hunyuan | 2026-05 | 7.8 Mrd. | Dicht | GQA | im_end_id im_newline_id im_start_id |
| Step-3.7-Flash | StepFun | 2026-05 | 198 Mrd. | Sparsames MoE | MHA | im_end_token im_patch_token im_start_token +5 |
| MiniMax-M3 | MiniMax | 2026-06 | 443.4 Mrd. | Sparsames MoE | GQA | image_grid_pinpoints image_seq_length image_token_index +3 |
| MiniMax-M3 MXFP8 | MiniMax | 2026-06 | 443.4 Mrd. | Sparsames MoE | GQA | image_grid_pinpoints image_seq_length image_token_index +3 |
| Kimi K3 | Kimi | 2026-07 | 2,780 Mrd. | Sparsames MoE | MLA | image_placeholder |
| GLM-5.3-Flash | GLM | 2026-08 | 328.9 Mrd. | Sparsames MoE | MLA | image_token_id image_start_token_id image_end_token_id |
| Qwen3.8 27B | Qwen | 2026-08 | 23.7 Mrd. | Dicht | GQA | image_token_id vision_end_token_id vision_start_token_id |
| Qwen3.8-Flash-Next | Qwen | 2026-08 | 123.6 Mrd. | Sparsames MoE | GQA | image_token_id vision_end_token_id vision_start_token_id |
Die vollständige Feldliste steht auf jeder Modellseite; die Tabelle zeigt die ersten Felder als Beleg.