Beijing-based Moonshot AI has released Kimi K3, a 2.8 trillion parameter model that the company describes in its technical blog as the world's first open 3T-class system and the largest open-weight AI model to date. Moonshot said K3 still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance, but it outperformed every other model in the company's evaluation suite, including Claude Opus 4.8 and GPT 5.5, across coding and agentic benchmarks. The model has a 1 million token context window, native vision, and activates just 16 of its 896 experts per token, roughly 1.8% of the pool. Full weights are due by July 27.Go deeper with TH Premium: AI and data centers(Image credit: Microsoft)Photonics and high-speed data movement is the next big AI bottleneckThe data center cooling state of playMassive AI data center buildouts are squeezing energy suppliesUltra Ethernet: The data center interconnection of tomorrowArena ranked K3 first in its Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing. API pricing is $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens. Kimi K2 launched a year ago at $0.60 per million input tokens, so uncached K3 input costs five times as much.Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.This is a 17-place jump from Kimi-k2.6 (#18 -> #1).In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics,… https://t.co/YDN3BufGkC pic.twitter.com/Oa6teaQnWpJuly 16, 2026Moonshot claims roughly a 2.5x improvement in scaling efficiency over Kimi K2, attributed to two architectural changes: Kimi Delta Attention, a hybrid linear attention scheme, and Attention Residuals, which change how information moves between layers. Quantization-aware training starts at the supervised fine-tuning stage, using MXFP4 weights and MXFP8 activations, a combination Moonshot says it chose for broad hardware compatibility. Bank of America analysts led by Alex Liu said in a note cited by CNBC that K3 shows large-scale pre-training plus architectural work can still deliver step-change gains for flagship Chinese models despite compute constraints.Moonshot's kernel optimization benchmark ran on Nvidia's H200, and what the blog identifies only as a "GPGPU from an alternative vendor," which the company didn't name. MiniTriton, a Triton-like compiler K3 built from scratch, is charted against Triton on an Nvidia L20, the cut-down Ada-based card sold into China under U.S. export rules. Moonshot recommends serving K3 on supernodes of 64 or more accelerators, keeping expert-parallel traffic inside one high-bandwidth domain. The blog doesn't say where the H200 hardware is; Congress passed a bill in January to close the offshore cloud rental loophole that gave Chinese firms remote access to restricted accelerators.In one case study, K3 spent a single 48-hour autonomous run designing a simulated inference chip for a nano model built on its own architecture, using open-source EDA tools and the Nangate 45nm library. The design closed timing at 100 MHz within 4mm squared, packed 1.46 million standard cells and an INT4 MAC array, and sustained more than 8,700 tokens per second of simulated decode.At the moment, every published K3 number is a claim made by Moonshot-reported or drawn from API access and can’t be verified until the weights are made public on July 27. Anthropic accused Moonshot in February of using 3.4 million Claude exchanges to train its models through distillation, and K3 now benchmarks within a few points of the models named in that complaint.