Modèles multimodaux

Les modèles à poids ouverts acceptant les images se multiplient. Cette page recense tous les modèles multimodaux du corpus et les preuves de config qui les identifient.

Critère (factuel, mécanique) : le config.json livré contient des champs liés à la vision — ids de tokens image/vision, réglages d’encodeur visuel ou champs de patchs d’image. Aucune autre déduction.

Modèle	Laboratoire	Publication	Paramètres	MoE	Attention	Preuve dans le config
Step-3	StepFun	2025-07	321 Md	MoE épars	MHA	`im_end_token` `im_patch_token` `im_start_token` +3
Hunyuan 7B	Hunyuan	2025-08	7.8 Md	Dense	GQA	`im_end_id` `im_newline_id` `im_start_id`
Kimi K2.5	Kimi	2026-02	1,040.2 Md	MoE épars	MLA	`use_unified_vision_chunk`
Qwen3.5 122B A10B	Qwen	2026-02	119.9 Md	MoE épars	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.5 27B	Qwen	2026-02	23.7 Md	Dense	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.5 35B A3B	Qwen	2026-02	33.9 Md	MoE épars	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.5 397B A17B	Qwen	2026-02	392.6 Md	MoE épars	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.5 9B	Qwen	2026-02	7.9 Md	Dense	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.5 9B Base	Qwen	2026-02	7.9 Md	Dense	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Kimi K2.6	Kimi	2026-04	1,040.2 Md	MoE épars	MLA	`use_unified_vision_chunk`
Qwen3.6 27B	Qwen	2026-04	23.7 Md	Dense	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.6 35B A3B	Qwen	2026-04	33.9 Md	MoE épars	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Hunyuan-TurboS 7B	Hunyuan	2026-05	7.8 Md	Dense	GQA	`im_end_id` `im_newline_id` `im_start_id`
Step-3.7-Flash	StepFun	2026-05	198 Md	MoE épars	MHA	`im_end_token` `im_patch_token` `im_start_token` +5
MiniMax-M3	MiniMax	2026-06	443.4 Md	MoE épars	GQA	`image_grid_pinpoints` `image_seq_length` `image_token_index` +3
MiniMax-M3 MXFP8	MiniMax	2026-06	443.4 Md	MoE épars	GQA	`image_grid_pinpoints` `image_seq_length` `image_token_index` +3
Kimi K3	Kimi	2026-07	2,780 Md	MoE épars	MLA	`image_placeholder`
GLM-5.3-Flash	GLM	2026-08	328.9 Md	MoE épars	MLA	`image_token_id` `image_start_token_id` `image_end_token_id`
Qwen3.8 27B	Qwen	2026-08	23.7 Md	Dense	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`
Qwen3.8-Flash-Next	Qwen	2026-08	123.6 Md	MoE épars	GQA	`image_token_id` `vision_end_token_id` `vision_start_token_id`

La liste complète des champs figure sur chaque page de modèle ; le tableau montre les premiers champs à titre de preuve.