Moondream: Moondream 3 Preview

{"about":"### Moondream 3 預覽版\n\nMoondream 3 Preview 屬於 Moondream 模型家族，這個家族以影像理解及場景中特定物件相關任務為核心。第三代改用專家混合架構：模型總計含 90 億參數，每個 token 使用約 20 億參數。這種分工結合了整體權重容量與處理每個 token 時較小的啟用部分。\n\nPreview 官方模型卡描述了 24 層、以 SigLIP 為基礎的視覺編碼器，以及 SuperBPE tokenizer。原始模型的專門技能包括影像問答、產生圖說、指出座標點，以及以矩形框選物件。這些是不同任務：解釋場景內容、描述整張影像，以及將物件對應至空間位置，會產生不同形式的結果。\n\n### Ailoot 的文字使用方式\n\nAiloot 為此模型提供文字輸入與文字答案。目前模型卡的介面不包含影像、物件座標或專用視覺方法。不過，開發商另為 Preview 提供了不含影像的提問範例：模型依一般文字指示解釋概念。\n\n此模式可用於說明、根據提供資訊回答問題，以及將文字整理成指定結構。例如，提供描述草稿、註明讀者，要求在保留事實的前提下改寫得更清楚。另一種做法是附上文字操作說明，要求改成行動順序，並為每一步提供簡短解釋。\n\n### 如何表達請求\n\n請直接說明任務與答案格式：簡短說明、連貫文章、清單或表格。若要分析特定資料，請將內容放入訊息，並註明哪些資訊必須維持原樣。若要在文字模式中詢問視覺場景，需要提供文字描述；模型處理的是所提供的描述，而非影像本身。\n\nAiloot 為 Moondream 3 Preview 標示 32,000 個 token 的上下文，包含目前討論的訊息與資料。Preview 名稱表示它是第三代預覽版本；此版本與 3.1 的特性應分開看待，尤其涉及原始模型的專門能力時。","alias_of":null,"aliases":[],"author":"Moondream","canonical_model":"moondream/moondream3-preview","ctx":null,"formats":["chat","gemini","messages","responses"],"id":"moondream/moondream3-preview","input":["text"],"kind":"text","max_out":null,"methods":[{"cache_read":null,"cache_write":null,"cache_write_1h":null,"cache_write_5m":null,"conditional_meter":null,"ctx":null,"family":"anthropic","fast_multiplier":1,"flex_multiplier":1,"fmt":"messages","http":"POST","img_in":null,"img_out":null,"in":0.6,"input":["text"],"label":"Anthropic Messages","max_out":null,"max_tool_calls":null,"out":5.0,"output":["text"],"path":"/v1/messages","request_model_id":"moondream/moondream3-preview","route_label":"Ultra","status_reads":[],"tiers":[],"tool_prices":null,"unit":"token","units":[]},{"cache_read":null,"cache_write":null,"cache_write_1h":null,"cache_write_5m":null,"conditional_meter":null,"ctx":null,"family":"openai","fast_multiplier":1,"flex_multiplier":1,"fmt":"responses","http":"POST","img_in":null,"img_out":null,"in":0.6,"input":["text"],"label":"OpenAI Responses","max_out":null,"max_tool_calls":null,"out":5.0,"output":["text"],"path":"/v1/responses","request_model_id":"moondream/moondream3-preview","route_label":"Ultra","status_reads":[],"tiers":[],"tool_prices":null,"unit":"token","units":[]},{"cache_read":null,"cache_write":null,"cache_write_1h":null,"cache_write_5m":null,"conditional_meter":null,"ctx":null,"family":"openai","fast_multiplier":1,"flex_multiplier":1,"fmt":"chat","http":"POST","img_in":null,"img_out":null,"in":0.6,"input":["text"],"label":"OpenAI Chat","max_out":null,"max_tool_calls":null,"out":5.0,"output":["text"],"path":"/v1/chat/completions","request_model_id":"moondream/moondream3-preview","route_label":"Ultra","status_reads":[],"tiers":[],"tool_prices":null,"unit":"token","units":[]},{"cache_read":null,"cache_write":null,"cache_write_1h":null,"cache_write_5m":null,"conditional_meter":null,"ctx":null,"family":"gemini","fast_multiplier":1,"flex_multiplier":1,"fmt":"gemini","http":"POST","img_in":null,"img_out":null,"in":0.6,"input":["text"],"label":"Gemini generateContent","max_out":null,"max_tool_calls":null,"out":5.0,"output":["text"],"path":"/v1beta/models/moondream/moondream3-preview:generateContent","request_model_id":"moondream/moondream3-preview","route_label":"Ultra","status_reads":[],"tiers":[],"tool_prices":null,"unit":"token","units":[]}],"model_aliases":[{"canonical_model_id":"moondream/moondream3-preview","request_model_id":"moondream/moondream3-preview","route_label":"Ultra"}],"name":"Moondream: Moondream 3 Preview","price":{"in":0.6,"in_label":null,"out":5.0,"out_label":null,"sort_in":0.6,"sort_out":5.0,"varies":false},"public_id":"moondream/moondream3-preview","short":"採用精簡 MoE 架構的 Moondream 3 預覽版；在 Ailoot 中回答文字問題並依指示處理文字。"}