A pure C inference engine runs the 2.78 trillion parameter Kimi K3 model on just 8GB RAM by streaming weights from disk, leveraging MoE sparsity (only 3.7% active per token) and computing directly on compressed formats, achieving correct output at 32.69 seconds per token on consumer hardware.
The article breaks down Kimi K3’s 2.8‑trillion‑parameter architecture, explains its 4‑bit MXFP4 quantization, calculates the ~1.4 TB memory requirement, shows that 8‑GPU clusters (e.g., NVIDIA B300 or AMD MI350X) are needed for self‑hosting, and compares these costs with the per‑token API pricing, highlighting when each option is economical.
A Bilibili creator released the open‑source Kimi K3 project, a pure front‑end Windows XP emulator that runs entirely in a web page without any VM or image download, faithfully reproducing the classic UI, built‑in applications, games, and even the original IE6 browsing experience.
The 47‑page Kimi K3 technical report, released on July 27, details the 2.8‑trillion‑parameter model’s novel LatentMoE, SiTU‑GLU, Quantile Balancing, Attention Residuals, and full‑stack NoPE design, explains how these solve activation‑explosion and load‑imbalance problems, and provides open‑source code for inference and agentic RL.
AirLLM introduces layer‑wise inference and per‑expert streaming to decouple VRAM usage from model size, enabling the 2.8‑trillion‑parameter Kimi K3 LLM to run on a single consumer‑grade GPU while preserving full‑precision accuracy and offering security‑focused insights.
A developer demonstrated that a 128 GB M5 Max Mac can stream‑load the 1.6 TB MXFP4 weights of the 2.8‑trillion‑parameter Kimi K3 model, achieving 0.32 token/s, while an 80‑GPU RTX 5090 cluster reaches 20 token/s, highlighting both feasibility and speed limits of large‑model inference on consumer hardware.
The article profiles the 401‑person core team behind the open‑source 2.8‑trillion‑parameter Kimi K3 model, detailing their academic backgrounds, landmark papers, engineering breakthroughs such as Mooncake KV‑Cache, MoBA, Muon optimizer, and the performance gains that let K3 run at only 38% of Claude Fable 5’s cost while boosting request capacity by over 75%.
Kimi K3, an open‑source LLM with 2.8 trillion parameters and a 57‑point intelligence score, outperforms many closed‑source rivals in benchmarks but suffers from a 40‑second first‑token delay and $0.72 per‑task cost, exposing the steep Test‑Time Compute hurdle that reshapes the AI market’s competitive landscape.
Kimi K3 scales to 2.8 T parameters, 104 B activation parameters and a 1 M‑token context by introducing KDA‑based compressed sequence state, AttnRes for depth‑wise residual selection, Stable LatentMoE for efficient expert routing, partial‑rollout reinforcement learning and Firecracker micro‑VM sandboxing, achieving roughly 2.5× scaling efficiency and notable benchmark gains over competing models.
The Kimi K3 model weights were released on HuggingFace (1.56 TB total), featuring mixed attention, Attention Residuals, and a Stable LatentMoE that together cut scaling cost by 2.5×, and a detailed cost‑benefit analysis shows how 80 consumer‑grade RTX 5090 cards can run the full 2.8‑trillion‑parameter model with 20 tok/s throughput, while highlighting memory‑saving quantization, KV‑cache design, and the steep price gap versus professional GPUs.
Kimi K3, a 2.8‑trillion‑parameter MoE model with a 100 k‑token context, has been fully open‑sourced along with its weights, technical report and infra (MoonEP, FlashKDA, AgentEnv), delivering programming and agent benchmark results that rival top closed models such as Claude Fable 5 and GPT‑5.6 while sparking debate over alleged distillation and emphasizing AI safety and open‑weight governance.
On July 27, Alibaba Cloud announced that its Lingjun Zhenwu M890 supernode instance has been successfully adapted to run the 2.8‑trillion‑parameter Kimi K3 model, achieving a 35 % reduction in first‑token latency, a 1.8× increase in decode throughput, and support for up to 1 M token context through joint chip, software‑stack and framework optimizations.
The article breaks down the massive hardware and budget requirements for running the open‑sourced 2.8‑trillion‑parameter Kimi K3 model locally, showing that a single node cannot hold the 1.56 TB weights and that realistic deployments start at ¥7‑8 million and can exceed ¥30 million for the recommended 64‑GPU supernode.
The article presents a detailed head‑to‑head assessment of the open‑source Kimi K3 model and the closed‑source GPT‑5.6 Sol, measuring their ability to generate playable 3D games, handle full‑stack development tasks, and comparing performance, token efficiency, and engineering completeness.
Jensen Huang’s first tweet highlighted a joint open‑weight AI letter, arguing that open‑source models like Kimi K3 are crucial for security, competition, and U.S. AI leadership, while also acknowledging the risks and policy actions needed to sustain an open ecosystem.
The author evaluates four large‑language models—Claude Fable 5, Qwen3.8‑Max, Kimi K3, and GPT‑5.6 Sol—by attempting a 1:1 front‑end recreation of Shopify’s Winter ’26 Editions page, and finds that Kimi K3 reproduces far more content and structure than Qwen3.8‑Max, though both have distinct strengths and weaknesses.
Security researcher Chaofan Shou revealed that Kimi K3’s AI model discovered a near zero‑click remote code execution flaw in Telegram’s desktop (Windows/macOS/Linux) and iOS clients, bypassing ASLR and requiring only a crafted data packet, with the exploit just one gadget away from full RCE.
Dean Ball, OpenAI’s new strategic‑future chief, praised Kimi K3’s 2.8‑trillion‑parameter performance but denounced its open‑weight release as a decelerationist move that could curb AI investment, spark regulatory panic, and reshape the US‑China AI competition landscape.
The article analyzes how Kimi K3’s strong performance on global leaderboards has reignited US debates over restricting Chinese open‑weight AI models, examining the model’s capabilities, pricing advantages, Microsoft’s evaluation, and the broader implications for US chip policy and AI market dynamics.
Kimi K3, the world’s first open‑source 2.8‑trillion‑parameter model, showcases novel attention and MoE techniques, scores near‑top on AI benchmarks, triggers valuation shifts for Anthropic, sparks debate among OpenAI leaders, and signals a broader industry move toward open‑source AI as DeepSeek V4 looms.
Less than a week after Kimi K3’s launch, soaring demand has maxed out Moonlight Dark’s GPU resources, forcing a pause on new and upgrade subscriptions, prompting a split of membership plans, driving six‑fold revenue growth, and accelerating a Hong‑Kong IPO within six months.
After Kimi K3’s official launch, the author evaluates its 2.8 T‑parameter, 1 M‑context, multimodal coding agent across three real‑world scenarios—a full‑stack hotspot‑tracking MVP, a Java project refactor fixing stock‑search encoding, and a 3A‑style game demo—detailing setup, performance, and limitations.
The author evaluates the newly released 2.8‑trillion‑parameter open‑source Kimi K3 model by having it generate a 3D rocket simulation, a 3D dinosaur runner game, a functional web‑based Excel, and an end‑to‑end Kaggle house‑price solution, revealing both impressive capabilities and notable limitations.
Latent MoE, demonstrated by NVIDIA’s Nemotron 3 Super and Moonshot AI’s 2.8 T‑parameter Kimi K3, compresses expert computations into a lower‑dimensional latent space, cutting memory reads and All‑to‑All traffic by fourfold, enabling more experts per token, higher accuracy, and up to 3.5× faster inference.
The article analyzes Kimi K3’s breakthrough in frontend generation, showing how it moves beyond merely writing code to grasping visual design, layout hierarchy, and iterative refinement, while offering practical prompts, task recommendations, and a realistic view of its strengths and risks.
Kimi K3, the world’s first open‑source 3‑trillion‑parameter LLM supporting 1 million‑token context and native visual understanding, tops the Arena.ai front‑end code benchmark, scores 57 on the AI Analysis Index, and introduces novel components such as KDA, Stable LatentMoE, and Quantile Balancing to achieve efficient scaling and strong cost‑performance.
On July 16, the Moon‑of‑Darkness team released Kimi K3, a 2.8‑trillion‑parameter open‑source large language model that introduces mixed‑linear attention, attention residuals, and a highly efficient Mixture‑of‑Experts design, achieving roughly 2.5× the scaling efficiency of its predecessor while approaching the performance of top closed‑source models.
The article reviews Kimi K3’s 2.8‑trillion‑parameter architecture, its performance advantages over previous models, pricing details, and demonstrates how the Claude Code Model Router (ccmr) enables seamless use of K3 within Claude Code through simple command‑line steps.
Kimi K3, the first open‑source 3‑trillion‑parameter model with 1 M context and native visual understanding, tops Arena.ai's front‑end code benchmark, scores 57 on the AI Analysis index, and introduces innovations such as KDA, Stable LatentMoE, Quantile Balancing, and Per‑Head Muon to achieve high training efficiency and competitive performance against closed models like Claude Fable 5 and GPT‑5.6 Sol.
Kimi K3, a newly released 2.8‑trillion‑parameter model with a 1‑million token context window, is fully open‑source and ranks third in overall AI intelligence scores, while achieving top‑three placements across a wide range of coding, agent, and multimodal benchmarks against leading models such as Claude Fable 5 and GPT‑5.6 Sol.
Kimi K3 arrives with 2.8 T parameters, native visual understanding, a 1 M‑token context window, novel KDA and Attention Residuals architecture, aggressive MoE sparsity, and pricing far below Western rivals, while achieving SOTA results on programming and agent benchmarks and demonstrating real‑world research and 3D‑game capabilities.
Kimi K3, a 2.8‑trillion‑parameter open‑source LLM, outperforms top closed‑source models in benchmarks, excels at long‑range coding, GPU kernel optimization, and multimodal tasks, while introducing novel attention mechanisms, a compact Triton‑like compiler, and even a prototype ASIC chip.
Kimi K3 is slated for a Q3 release with a massive 2.5 trillion parameters, surpassing DeepSeek V4 Pro and Baidu Wenxin 5.0, reigniting the large‑model arms race and prompting a debate between scale, efficiency, and ecosystem‑driven approaches.
