In less than six months, Moonshot AI's smart assistant has increased its capacity tenfold, now supporting 2 million characters of lossless context.

On March 18, 2024, general artificial intelligence startup Moonshot AI announced a breakthrough in large model long-context window technology: its Kimi intelligent assistant now supports lossless context of up to 2 million Chinese characters, opening product beta testing effective immediately.

From the first funding round onward, Monolith Management has invested in Moonshot AI for three consecutive rounds, consistently backing the company's bold innovation and technical breakthroughs. We supported and witnessed the initial launch of the Kimi intelligent assistant in October 2023. In less than six months, the team has increased its lossless context length by an order of magnitude. We believe this leap will unlock new application scenarios for users.

Users with needs for ultra-long lossless context capabilities can apply for early access on the Kimi intelligent assistant web version at kimi.ai.

Kimi is a conversational AI assistant product built by Moonshot AI on its self-developed hundred-billion-parameter large model. At its launch in October 2023, it supported approximately 200,000 Chinese characters of lossless context input — a record for consumer AI products. In November 2023, Kimi officially opened to the public. Its outstanding long-context processing capabilities helped users unlock many new use cases, including translating and understanding professional academic papers, assisting with legal analysis, organizing dozens of invoices at once, and quickly comprehending API development documentation, earning strong user reviews.

In less than six months, Moonshot AI increased Kimi's lossless context length by an order of magnitude, from 200,000 to 2 million characters. Because it did not follow the conventional gradual improvement path, the technical difficulty the Moonshot AI team encountered increased exponentially. To achieve better long-window lossless compression performance, Moonshot AI's R&D and technical teams redesigned and redeveloped from the ground up across model pre-training, alignment, and inference — eschewing technical shortcuts like "sliding windows" and "downsampling," overcoming numerous fundamental technical challenges to achieve this breakthrough.

This time, Moonshot AI offered a "modest proposal" showcasing several example use cases for ultra-long lossless context. For instance, users can upload hundreds of thousands of words of classic Texas Hold'em tutorials and have Kimi play the role of a poker expert to guide their playing strategy.

Upload a complete, nearly million-word traditional Chinese medicine diagnostic manual, and have Kimi provide diagnostic recommendations based on user queries.

Upload NVIDIA's complete financial reports from recent years, and have Kimi become a NVIDIA financial research expert, helping users analyze and summarize important historical development milestones.

Upload source code from a code repository, and you can ask Kimi about any detail of the codebase — even ancient code without any comments can help you quickly map out its structure.

Fields that once required 10,000 hours to master can now be approached at a junior expert level by Kimi in just 10 minutes. Users can discuss topics in the field with Kimi, have Kimi help them practice professional skills, or inspire new ideas. With Kimi supporting 2 million characters of lossless context, rapidly learning any new field becomes much easier.

Rapidly organizing large volumes of materials is a frequent workplace challenge. Now Kimi can intensively read 500 or even more files at once, helping users quickly analyze all their contents, with support for natural language information queries and filtering — dramatically improving information processing efficiency. For example, an HR professional can, based on business needs, quickly have Kimi identify candidates with experience in a specific industry who also graduated with computer science degrees from a recent batch of 500 resumes, screening and identifying suitable candidates far more efficiently.

From full-length novels, stories, or scripts, rediscovering subtle clues worth exploring and mining deep details is a passion for many film and TV entertainment IP enthusiasts. If you feed Kimi the complete Empresses in the Palace script of hundreds of thousands of words and ask which details indicate that Zhen Huan's child belongs to the Prince of Guo — Kimi can dig deep across different time periods and various scenes to unearth the emotional threads between Zhen Huan and the Prince of Guo and the truth about their child, comparable to a "Zhen" scholar who has watched the series dozens of times.

"We believe that the order-of-magnitude increase in large model lossless context length will further help everyone expand their imagination of AI application scenarios, including complete code repository analysis and understanding, autonomous agents capable of completing multi-step complex tasks, lifelong assistants that don't forget critical information, truly unified multimodal models, and more," said Xu Xinran, Moonshot AI's VP of Engineering. "Whether it's memory, compute, or network bandwidth, every historical upgrade in foundational technology has unlocked new product forms and application scenarios. We are full of anticipation for what innovative opportunities beyond imagination the 2-million-character lossless context Kimi might bring."

"On the path to artificial general intelligence, lossless long context will be a crucial foundational technology. From word2vec to RNN, LSTM, and then to Transformer, the essence of all historical model architecture evolution has been improving effective, lossless context length," Moonshot AI founder Zhilin Yang said in a previous interview. "Context length may have its own Moore's Law, but it only counts as meaningful scaling if you optimize both length and lossless compression level simultaneously."

Based on feedback from many Kimi users, the 200,000-character lossless long context helped them open up a new world of AI applications and delivered greater value. But as they attempted more complex tasks and interpreted longer documents, they still encountered situations where conversation length exceeded limits. This is a direct reason why large model products need to continue increasing their lossless context length.

Moreover, Kimi's intelligent search is inseparable from large model lossless long context capability. Multiple sources that Kimi actively retrieves become part of the context passed to the model for reasoning. It is precisely because Kimi's large model supports sufficiently long context windows with sufficiently low information loss that the Kimi intelligent assistant can output high-quality results and deliver a fundamentally different search experience — Kimi can proactively search the internet, analyze, and summarize the most relevant pages based on user queries, generating more direct and accurate answers. For example, users can have Kimi actively search for and compare the latest financial report data of two companies in the same sector, directly generating comparison tables and saving substantial research time. Traditional search engines typically can only return web links mixed with advertising based on user queries.

Another metric closely tied to large model lossless context capability is instruction following ability. Instruction following manifests in two main aspects: first, whether the model consistently follows user instructions and understands user needs across multi-turn conversations; second, whether the model can follow complex instructions that may sometimes run thousands or tens of thousands of words. Based on user feedback since product launch, Kimi's multi-turn interaction and ultra-long instruction following capabilities also demonstrate significant advantages.

With daily model capability upgrades and the rollout of iOS apps, Android apps, mini-programs, Web, and other multi-platform access, the Kimi intelligent assistant has become an indispensable AI assistant for more and more users in their work and lives. Following the March 18 launch of 2-million-character ultra-long context beta applications, Moonshot AI will gradually open access to more users to experience the Kimi intelligent assistant with ultra-long lossless context capabilities, looking forward to co-creating intelligence with more users.