AI Plugin from Sakana AI Simplifies Document Processing for Large Models
Large language models' struggle with processing long texts, often called "memory anxiety," may soon be a thing of the past. Recently, Tokyo-based AI startup Sakana AI unveiled two breakthrough technologies: Text-to-LoRA (T2L) and Doc-to-LoRA (D2L). Utilizing an innovative "super network" architecture, these technologies enable large models to "absorb" massive documents or learn new tasks in under a second, all without retraining.

AI developers have long faced a difficult trade-off: cramming lengthy documents into a chat window—which slows responses and consumes high memory—or paying the high cost of fine-tuning a model. Sakana AI introduces a third option. Through a one-time payment for pre-training, it generates minuscule weight plugins (LoRA) to achieve low-cost, efficient model adaptation.
Doc-to-LoRA: Memory Needs Drop from 12GB to Just 50MB
This represents the most impressive technology in the latest release. Processing a 128,000-token document (roughly 100,000 words) with traditional methods demands over 12GB of VRAM to store information. With D2L, the model can directly "digest" this information into a plugin smaller than 50MB.
Remarkable Speed: Legacy technology requires 40 to 100 seconds to process a document, while D2L accomplishes it in under 1 second.
Pushing Boundaries: It allows models to handle text up to four times longer than their original context window, achieving near-perfect accuracy in "needle-in-a-haystack" retrieval tests.
Text-to-LoRA: Tailor AI with Everyday Language
Text-to-LoRA makes models more responsive. Users simply describe a task in natural language—for example, "help me solve a complex math competition problem"—and the system automatically generates a dedicated plugin to enhance performance. Experiments reveal that adapters created this way can outperform dedicated models trained from scratch on math and logical reasoning tasks.
Powerful Cross-Modal Tech: Enabling Text Models to "See" Images
Researchers discovered an unexpected bonus: D2L exhibits strong cross-modal abilities. By mapping visual information into the parameters of a pure text model, a model that has never processed images before can classify them with an accuracy of **75.03%**.
Sakana AI’s accomplishments significantly lower the barrier for individuals and companies to customize private AI models. They also pave a new way toward developing more lightweight and intelligent general artificial intelligence (AGI).
Paper: https://arxiv.org/pdf/2602.15902
Related article
Anthropic hit with US export controls amid long-running criticisms of its own products
Business Insider reports that Anthropic, a leading AI firm, faces backlash for repeatedly highlighting risks associated with its own technology. CEO Dario Amodei recently published articles warning that the company’s newest model presents significant
Tencent Unveils BrowserSkill, an Open-Source Browser Automation Tool Empowering AI Agents
As artificial intelligence technology accelerates its practical application, enabling AI Agents to efficiently manage daily web tasks has become a key priority for developers. Recently, Tencent open-sourced BrowserSkill, a browser automation tool spe
ByteDance reshapes AI strategy as large model ARR hits $4 billion amid integration with collaboration tools and core models
On July 30, ByteDance unveiled a significant restructuring of its AI division, splitting the Feishu collaboration platform team and integrating its components into the Doubao and Volcano Engine units. This strategic shift is designed to enhance the i
Related Special Topic Recommendations
Comments (0)
0/500
Large language models' struggle with processing long texts, often called "memory anxiety," may soon be a thing of the past. Recently, Tokyo-based AI startup Sakana AI unveiled two breakthrough technologies: Text-to-LoRA (T2L) and Doc-to-LoRA (D2L). Utilizing an innovative "super network" architecture, these technologies enable large models to "absorb" massive documents or learn new tasks in under a second, all without retraining.

AI developers have long faced a difficult trade-off: cramming lengthy documents into a chat window—which slows responses and consumes high memory—or paying the high cost of fine-tuning a model. Sakana AI introduces a third option. Through a one-time payment for pre-training, it generates minuscule weight plugins (LoRA) to achieve low-cost, efficient model adaptation.
Doc-to-LoRA: Memory Needs Drop from 12GB to Just 50MB
This represents the most impressive technology in the latest release. Processing a 128,000-token document (roughly 100,000 words) with traditional methods demands over 12GB of VRAM to store information. With D2L, the model can directly "digest" this information into a plugin smaller than 50MB.
Remarkable Speed: Legacy technology requires 40 to 100 seconds to process a document, while D2L accomplishes it in under 1 second.
Pushing Boundaries: It allows models to handle text up to four times longer than their original context window, achieving near-perfect accuracy in "needle-in-a-haystack" retrieval tests.
Text-to-LoRA: Tailor AI with Everyday Language
Text-to-LoRA makes models more responsive. Users simply describe a task in natural language—for example, "help me solve a complex math competition problem"—and the system automatically generates a dedicated plugin to enhance performance. Experiments reveal that adapters created this way can outperform dedicated models trained from scratch on math and logical reasoning tasks.
Powerful Cross-Modal Tech: Enabling Text Models to "See" Images
Researchers discovered an unexpected bonus: D2L exhibits strong cross-modal abilities. By mapping visual information into the parameters of a pure text model, a model that has never processed images before can classify them with an accuracy of **75.03%**.
Sakana AI’s accomplishments significantly lower the barrier for individuals and companies to customize private AI models. They also pave a new way toward developing more lightweight and intelligent general artificial intelligence (AGI).
Paper: https://arxiv.org/pdf/2602.15902
Anthropic hit with US export controls amid long-running criticisms of its own products
Business Insider reports that Anthropic, a leading AI firm, faces backlash for repeatedly highlighting risks associated with its own technology. CEO Dario Amodei recently published articles warning that the company’s newest model presents significant
Tencent Unveils BrowserSkill, an Open-Source Browser Automation Tool Empowering AI Agents
As artificial intelligence technology accelerates its practical application, enabling AI Agents to efficiently manage daily web tasks has become a key priority for developers. Recently, Tencent open-sourced BrowserSkill, a browser automation tool spe
ByteDance reshapes AI strategy as large model ARR hits $4 billion amid integration with collaboration tools and core models
On July 30, ByteDance unveiled a significant restructuring of its AI division, splitting the Feishu collaboration platform team and integrating its components into the Doubao and Volcano Engine units. This strategic shift is designed to enhance the i





Home






