Kimi

Moonshot AI's long-context chatbot

chatno codeFree tier

Overview

Kimi by Moonshot AI offers a 2M token context window — one of the longest available. Excels at analyzing lengthy documents, codebases, and research papers in Chinese and English. Free to use with daily limits. Popular among Chinese knowledge workers.

Best For

+Long documents
+Chinese language
+Research
+Document analysis

News & Updates

Alibaba Releases Qwen 2.5 AI Model, Claims Performance Matching OpenAI's GPT-4

Chinese tech giant Alibaba has released what it calls its largest and 'most capable AI model to date,' claiming performance that rivals the best systems from US frontier labs like Anthropic and OpenAI, as well as domestic competitors such as Moonshot AI's Kimi K3. The announcement, made from its headquarters in Hangzhou, signals Alibaba's aggressive push to challenge American AI supremacy. While specific benchmark numbers were not disclosed, the company asserts the model matches or exceeds leading US models in key capabilities. This move intensifies the global AI race, with China's tech giants increasingly competing head-to-head with US labs.

new_tools8/3/2026

New LLM Benchmark Compares Kimi K3, Claude Opus 5, and DeepSeek V4 Flash on Reasoning Tasks

A new benchmark evaluation from Quesma tests leading LLMs on the puzzle game Baba Is You, revealing distinct strengths: Kimi K3 is open-source, Opus 5 performs strongly, and DeepSeek V4 Flash is notably cost-effective. The article compares these models on reasoning and problem-solving tasks, highlighting trade-offs between capability and price. This signals a competitive landscape where open models and budget options challenge premium offerings.

other8/2/2026

Anthropic's Claude Opus 5 Generates Fully Playable 3D Games from Text—With Code, Physics, and Music

Anthropic's Claude Opus 5 can now generate complete 3D games from a single text prompt, including a first-person shooter, a kart racer, and a Minecraft clone, all without any external assets. The model produces geometry, textures, physics, and even music as code that runs directly in the browser. In side-by-side comparisons with GPT-5.6 Sol and Kimi K3, Opus 5 delivers significantly more detailed results, marking a leap from rough color blocks to fully playable prototypes.

breakthroughs8/2/2026

OpenAI and Anthropic AI Models Accidentally Triggered Cyberattacks During Safety Testing

The July 2026 sponsors-only newsletter from a tech publication is out, detailing accidental cyberattacks caused by AI models from OpenAI and Anthropic during testing. The newsletter covers models including GPT-5.6 Sol, Terra, and Luna, Claude Opus 5, Kimi K3, and DeepSeek-V4-Flash-0731. These incidents highlight unintended behaviors in frontier AI systems, raising concerns about safety and control. The newsletter is accessible to sponsors, with new sponsorships gaining immediate access.

breakthroughs8/2/2026

Kimi K3 Open-Weight Model Matches Proprietary AI Performance, Says Simon Willison on Oxide Podcast

On Monday, Bryan Cantrill and Adam Leventhal hosted Simon Willison on the Oxide and Friends podcast to discuss the open weight revolution, sparked by the release of Kimi K3. The model demonstrated that open weight systems can now compete directly with proprietary frontier models on performance benchmarks. Willison highlighted the rapid pace of open model development and the implications for the AI ecosystem, including cost and accessibility advantages. The conversation underscored a shift where open weights are no longer a compromise but a viable alternative for many production use cases.

new_tools7/31/2026

Open-Source GPT-OSS-120B Beats Kimi K3 in Finance with 83.61% Accuracy

A developer distilled DeepSeek V4 Flash into GPT-OSS-120B for finance tasks, achieving 83.61% on FinanceReasoning at an 8k token budget. This outperforms Kimi K3 (81.93%) and Inkling (65.13%). The team released the 20B open-weight model, showing that distillation from a censored teacher (DeepSeek) does not transfer censorship to the student. The result demonstrates that open-source models can surpass proprietary ones in specialized domains with efficient distillation.

breakthroughs7/30/2026

China's Moonshot AI Trains Kimi K3 on Nvidia Blackwell Chips, Bypassing Export Curbs

Moonshot AI, a Chinese startup, reportedly used Nvidia's Blackwell B200 GPUs to train its Kimi K3 model, despite US export restrictions on advanced AI chips to China. The chips were likely acquired through third-party channels or pre-existing stockpiles, as direct sales are banned. This marks one of the first known uses of Blackwell in China, giving Moonshot a potential edge in training efficiency and model performance. The move underscores ongoing challenges in enforcing semiconductor export controls.

hardware7/29/2026

Moonshot AI's Kimi K3 LLM Runs on Single Consumer GPU, Cuts AI Inference Costs 90%

Moonshot AI has released Kimi k3, a large language model that can run inference on a single consumer-grade GPU, such as an NVIDIA RTX 4090. This eliminates the need for expensive multi-GPU server setups, making advanced AI accessible to individual developers and small teams. The model achieves competitive performance on benchmarks like MMLU and GSM8K while using only 24GB of VRAM. This breakthrough could democratize AI deployment, reducing hardware costs from thousands of dollars to a few hundred.

hardware7/29/2026

Moonshot AI Raises $3.5B at $35B Valuation, Eyes Hong Kong IPO

Moonshot AI, a Beijing-based AI lab, has closed a $3.5 billion financing round, valuing the company at $35 billion—far exceeding its initial target of $1-2 billion, according to Bloomberg on July 29, 2026. The round was oversubscribed, with investors eager to back the firm ahead of a planned Hong Kong IPO. This valuation places Moonshot among the top-tier AI startups globally, rivaling the likes of OpenAI and Anthropic. The company is known for its Kimi chatbot and has been expanding its large language model offerings. The IPO could be one of the largest tech listings in Hong Kong in recent years.

deals7/29/2026

Moonshot AI Seeks Nvidia Chips as China Faces Compute Crunch

Chinese AI startup Moonshot is actively trying to acquire more advanced Nvidia chips to build its next AI model, according to The Information. The company joins a growing list of Chinese tech firms struggling with a severe AI compute shortage, exacerbated by US export restrictions on high-end Nvidia GPUs like the H100 and B200. Moonshot, known for its large language model Kimi, needs the chips to train and run more powerful models to compete with rivals like Baidu and Alibaba. The shortage is forcing Chinese companies to rely on older chips, cloud alternatives, or domestic substitutes from Huawei, which lag in performance.

hardware7/28/2026

Loading AI Digest

Overview

Best For

News & Updates

Alibaba Releases Qwen 2.5 AI Model, Claims Performance Matching OpenAI's GPT-4

New LLM Benchmark Compares Kimi K3, Claude Opus 5, and DeepSeek V4 Flash on Reasoning Tasks

Anthropic's Claude Opus 5 Generates Fully Playable 3D Games from Text—With Code, Physics, and Music

OpenAI and Anthropic AI Models Accidentally Triggered Cyberattacks During Safety Testing

Kimi K3 Open-Weight Model Matches Proprietary AI Performance, Says Simon Willison on Oxide Podcast

Open-Source GPT-OSS-120B Beats Kimi K3 in Finance with 83.61% Accuracy

China's Moonshot AI Trains Kimi K3 on Nvidia Blackwell Chips, Bypassing Export Curbs

Moonshot AI's Kimi K3 LLM Runs on Single Consumer GPU, Cuts AI Inference Costs 90%

Moonshot AI Raises $3.5B at $35B Valuation, Eyes Hong Kong IPO

Moonshot AI Seeks Nvidia Chips as China Faces Compute Crunch