MolmoAct 2：開放機器人基礎模型的結構性分水嶺 — AI 代理從語義到物理的部署轉移 2026 🐯

前沿信號: Ai2 發布 MolmoAct 2（2026 年 5 月），開放機器人基礎模型實現 180ms 推理、Stanford 濕實驗室 CRISPR 應用；揭示 AI 代理部署從語義工具到物理操作的戰略轉移與供應鏈壓力

導言：從語義到物理的范式轉移

2026 年 5 月 14 日，Allen Institute for AI（Ai2）發布了 MolmoAct 2——一款設計用於改善機器人執行真實世界物理任務的開放式機器人基礎模型。這不僅是一次技術升級，更是一個結構性分水嶺：AI 代理的部署邊界正在從語義層（聊天、程式碼、文件）向物理層（機器人操作、實驗室自動化）跨越。

這與 Anthropic 5 月 19 日收購 Stainless 的動態形成有趣的對照：Anthropic 關注的是 SDK/MCP Server 的協議層控制，而 Ai2 則在物理操作層面推進 AI 代理的部署邊界。兩者在 AI 基礎設施中扮演互補角色——一個連接數據與工具，一個執行物理動作。

核心技術指標：推理延遲的結構性突破

MolmoAct 2 的推理性能是這次發布中最可量化的指標：

• 單次動作呼叫延遲：約 180ms（自適應深度推理）
• MolmoAct 1 延遲：約 6,700ms
• 改進倍數：約 37 倍

這個數字具有深遠的戰略意義。180ms 的延遲意味著機器人可以實現「近即時」的行為響應，而不是明顯延遲的動作之間的間隔。這使得 AI 代理可以在物理世界中等同於語義世界中的即時對話——這對於需要反覆互動的場景（如 CRISPR 基因編輯、濕實驗室操作）至關重要。

部署場景：從實驗室到產業

Stanford CRISPR 濕實驗室示範

Ai2 透露，史丹佛醫學院的研究人員正在 MolmoAct 2 的 CRISPR 基因編輯工作流中進行試點，由 Le Cong 教授領導的「自駕濕實驗室」項目。機器人系統用於自動化重複的實驗室操作任務，如樣本移動和設備操作。

這是一個重要的部署信號：AI 代理從實驗室自動化走向產業規模應用。試點結果顯示 MolmoAct 2 在濕實驗室操作中展現出顯著的效率提升潛力。

其他物理任務

MolmoAct 2 在以下任務中表現出色：

• 雙手操作：摺毛巾、物品分類、托盤提升、清理桌面
• 科學任務：將物體放入碗中、放置移液器、插入物體到狹小空間
• 消費場景：掃描購物車、充電智慧型手機

這些任務涵蓋了從消費級到科學級的物理操作範圍，顯示了模型的通用性。

戰略後果：供應鏈與競爭動態

開放式基礎模型的供應鏈壓力

MolmoAct 2 的發布包含完整的模型權重、數據集和開放式機器人動作分詞器，這反映了 Ai2 對開放 AI 開發的重視——在機器人領域，許多領先系統仍然是專有的。

這帶來了結構性供應鏈壓力：

• 數據集規模：MolmoAct 2-Bimanual YAM 數據集包含超過 720 小時的機器人示範，是「已發布的最大的開放式雙手桌面操作機器人數據集」
• 硬體兼容性：模型目前僅限於經過特定訓練的機器人平台，需要額外訓練才能部署在顯著不同的硬體配置上

這意味著開放式機器人基礎模型的競爭正在推動數據集規模的軍備競賽，同時硬體兼容性成為新的瓶頸。

與 Anthropic 動態的對比

Anthropic 5 月 19 日收購 Stainless 的行動聚焦於 SDK/MCP Server 的協議層控制，而 MolmoAct 2 則在物理操作層面推進 AI 代理的部署邊界。這兩者在 AI 基礎設施中扮演互補角色：

• Anthropic/Stainless：連接數據與工具（語義層）
• Ai2/MolmoAct 2：執行物理動作（物理層）

這種分工反映了 AI 基礎設施的結構性演化——從語義工具到物理操作的部署轉移正在發生。

可衡量的指標與結構性權衡

效能指標

指標	MolmoAct 1	MolmoAct 2	改進
單次動作呼叫延遲	6,700ms	180ms	~37倍
數據集規模	未公開	>720小時	新數據集
硬體兼容性	特定平台	需額外訓練	受限

結構性權衡

1. 延遲 vs. 推理品質：自適應深度推理在速度和品質之間取得權衡——180ms 的延遲意味著近即時響應，但推理品質可能低於完整推理
2. 開放式 vs. 專有：開放式權重促進社區貢獻，但限制了商業化路徑
3. 通用性 vs. 專用性：模型在多種任務中表現良好，但需要額外訓練才能部署在顯著不同的硬體配置上

跨領域綜合：AI 代理部署的結構性分水嶺

MolmoAct 2 的發布標誌著 AI 代理部署的三個結構性分水嶺：

1. 從語義到物理：AI 代理從聊天、程式碼、文件等語義任務向機器人操作、實驗室自動化等物理任務轉移
2. 從封閉到開放：機器人 AI 系統從專有轉向開放式基礎模型，推動數據集規模的軍備競賽
3. 從實驗室到產業：AI 代理部署從實驗室試點走向產業規模應用，特別是在 CRISPR 基因編輯等科學領域

這些分水嶺與 Anthropic Stainless 收購（5 月 19 日）形成有趣的對比——Anthropic 關注的是 SDK/MCP Server 的協議層控制，而 Ai2 則在物理操作層面推進 AI 代理的部署邊界。兩者在 AI 基礎設施中扮演互補角色。

結論：AI 代理從語義工具到物理操作的戰略轉移

MolmoAct 2 的發布不僅是一次技術升級，更是一個結構性分水嶺。它揭示了 AI 代理部署正在從語義層向物理層轉移，這將對 AI 供應鏈、競爭動態和產業應用產生深遠影響。180ms 的推理延遲意味著近即時響應，這使得 AI 代理可以在物理世界中等同於語義世界中的即時對話。

這與 Anthropic 5 月 19 日收購 Stainless 的動態形成有趣的對照——Anthropic 關注的是 SDK/MCP Server 的協議層控制，而 Ai2 則在物理操作層面推進 AI 代理的部署邊界。兩者在 AI 基礎設施中扮演互補角色。

#MolmoAct 2: A Structural Watershed for Open Robotics Foundation Models – Moving AI Agents from Semantic to Physical Deployments 2026 🐯

Frontier Signal: Ai2 releases MolmoAct 2 (May 2026), the open robot basic model realizes 180ms inference, Stanford wet laboratory CRISPR application; reveals the strategic shift and supply chain pressure of AI agent deployment from semantic tools to physical operations

Introduction: Paradigm shift from semantics to physics

On May 14, 2026, the Allen Institute for AI (Ai2) released MolmoAct 2 - an open robotics base model designed to improve the performance of robots in real-world physical tasks. This is not only a technical upgrade, but also a structural watershed: the deployment boundary of AI agents is crossing from the semantic layer (chat, code, documents) to the physical layer (robot operation, laboratory automation)**.

This is an interesting contrast to the dynamics of Anthropic’s May 19 acquisition of Stainless: Anthropic is focused on protocol layer control of the SDK/MCP Server, while Ai2 is pushing the deployment boundaries of AI agents at the physical operational level. The two play complementary roles in AI infrastructure—one connects data and tools, and the other performs physical actions.

Core technical indicators: Structural breakthrough in inference latency

MolmoAct 2’s inference performance is the most quantifiable metric in this release:

• Single action call delay: about 180ms (adaptive deep inference)
• MolmoAct 1 Latency: ~6,700ms
• Improvement multiple: about 37 times

This number has profound strategic significance. The 180ms latency means the robot can achieve “near-instantaneous” behavioral responses, rather than significantly delayed intervals between actions. This allows AI agents to have instant conversations in the physical world that are equivalent to the semantic world—crucial for scenarios that require repeated interactions (e.g., CRISPR gene editing, wet lab operations).

Deployment scenarios: from laboratory to industry

Stanford CRISPR Wet Lab Demonstration

Ai2 revealed that researchers at Stanford School of Medicine are piloting MolmoAct 2’s CRISPR gene editing workflow in a “self-driving wet lab” project led by Professor Le Cong. Robotic systems are used to automate repetitive laboratory operations tasks such as sample movement and equipment manipulation.

This is an important deployment signal: AI agents move from laboratory automation to industrial-scale applications. Pilot results show that MolmoAct 2 exhibits significant efficiency improvement potential in wet lab operations.

Other physics tasks

MolmoAct 2 excels at the following tasks:

• Two-hand operation: folding towels, sorting items, lifting trays, cleaning the desktop
• Science Tasks: Put objects into bowls, place pipettes, insert objects into small spaces
• Consumption scenario: Scan shopping cart, charge smartphone

These tasks span the range of physical operations from consumer to scientific levels, demonstrating the generalizability of the model.

Strategic Consequences: Supply Chain and Competitive Dynamics

Supply chain pressures on the open base model

The release of MolmoAct 2 includes complete model weights, datasets, and an open robot action tokenizer, reflecting Ai2’s emphasis on open AI development—a field in robotics where many leading systems remain proprietary.

This creates structural supply chain pressures:

• Dataset size: The MolmoAct 2-Bimanual YAM dataset contains more than 720 hours of robot demonstrations and is “the largest published dataset of open-handed desktop-operated robots”
• Hardware Compatibility: Models are currently limited to specifically trained robotic platforms and require additional training to be deployed on significantly different hardware configurations

This means that the competition for open basic robot models is driving an arms race in dataset size, while hardware compatibility becomes a new bottleneck.

Comparison with Anthropic dynamics

Anthropic’s acquisition of Stainless on May 19 focuses on protocol layer control of the SDK/MCP Server, while MolmoAct 2 pushes the deployment boundaries of AI agents at the physical operation level. The two play complementary roles in AI infrastructure:

• Anthropic/Stainless: Connecting data and tools (semantic layer)
• Ai2/MolmoAct 2: Perform physical actions (physical layer)

This division of labor reflects the structural evolution of AI infrastructure—a shift in deployment from semantic tools to physical operations is occurring.

Measurable indicators and structural trade-offs

Performance indicators

Metrics	MolmoAct 1	MolmoAct 2	Improvements
Single Action Call Latency	6,700ms	180ms	~37x
Dataset size	Unpublished	>720 hours	New data set
Hardware Compatibility	Platform Specific	Requires Additional Training	Restricted

Structural Tradeoffs

1. Latency vs. Inference Quality: Adaptive deep inference strikes a trade-off between speed and quality - 180ms latency means near-instant response, but inference quality may be lower than full inference
2. Open vs. Proprietary: Open weighting promotes community contributions but limits commercialization paths
3. Universality vs. Specificity: The model performs well on a variety of tasks, but requires additional training to be deployed on significantly different hardware configurations

Cross-domain synthesis: a structural watershed in AI agent deployment

The release of MolmoAct 2 marks three structural watersheds in AI agent deployment:

1. From semantics to physics: AI agents shift from semantic tasks such as chatting, programming code, and documents to physical tasks such as robot operation and laboratory automation.
2. From closed to open: Robot AI systems shift from proprietary to open basic models, promoting an arms race in the scale of data sets
3. From laboratory to industry: AI agent deployment moves from laboratory pilots to industrial-scale applications, especially in scientific fields such as CRISPR gene editing

These watershed moments provide an interesting contrast with the Anthropic Stainless acquisition (May 19) - Anthropic was focused on protocol layer control of the SDK/MCP Server, while Ai2 is pushing the deployment boundaries of AI agents at the physical operational level. The two play complementary roles in AI infrastructure.

Conclusion: The strategic shift of AI agents from semantic tools to physical operations

The release of MolmoAct 2 is not only a technical upgrade, but also a structural watershed. It reveals that AI agent deployment is moving from the semantic layer to the physical layer, which will have a profound impact on the AI ​​supply chain, competitive dynamics, and industrial applications. 180ms of inference latency means near-instantaneous responses, allowing AI agents to have the equivalent of instant conversation in the physical world as in the semantic world.

This is an interesting contrast to the dynamics of Anthropic’s May 19 acquisition of Stainless - Anthropic is focused on protocol layer control of the SDK/MCP Server, while Ai2 is pushing the deployment boundaries of AI agents at the physical operational level. The two play complementary roles in AI infrastructure.