Moonshot AI has introduced Kimi K3, a 2.8 trillion-parameter model the Beijing company says is the first open 3T-class system and the largest open-weight AI model so far. The claim matters for developers because Moonshot is pitching K3 as a high-end coding and agentic model that can be run outside a closed commercial API, once the weights arrive.
According to Moonshot’s technical blog, the model trails Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol on broad overall performance. Moonshot says K3 still beat every other model in its evaluation set, including Claude Opus 4.8 and GPT 5.5, on coding and agentic benchmarks.
The company says K3 has a 1 million-token context window and native vision support. It uses a mixture-of-experts design with 896 experts, but activates 16 experts for each token, about 1.8 percent of the pool. Moonshot says the full weights are scheduled for release by July 27.
Coding benchmark claims
Arena ranked K3 first in its Frontend Code evaluation with 1,679 points, ahead of Claude Fable 5, in blind testing by developers. A post about the result said K3 rose from 18th place for Kimi K2.6 to first place, and ranked first in six of seven frontend domains.
Moonshot’s API price for K3 is $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens. Kimi K2, released a year earlier, cost $0.60 per million input tokens, so K3’s uncached input price is five times higher than K2’s listed input price.
Architecture and hardware
Moonshot attributes K3’s gains to about a 2.5 times improvement in scaling efficiency over Kimi K2. The company points to two design changes: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which alter how information passes between layers.
The company says it begins quantization-aware training during supervised fine-tuning, using MXFP4 weights and MXFP8 activations. Moonshot says that format mix was chosen for broad hardware compatibility.
Bank of America analysts led by Alex Liu, in a note cited by CNBC, said K3 suggests large-scale pretraining and architecture changes can still produce major gains for Chinese flagship models despite compute limits.
Moonshot’s kernel optimization benchmark used Nvidia’s H200 and what the company described only as a GPGPU from an alternative vendor. The blog does not name that vendor. Moonshot also charted MiniTriton, its Triton-like compiler built from scratch, against Triton on Nvidia’s L20, the reduced Ada-based card sold into China under U.S. export controls.
For serving K3, Moonshot recommends supernodes with 64 or more accelerators, keeping expert-parallel traffic inside a high-bandwidth domain. The company did not say where the H200 systems used in its benchmark were located. Congress passed a bill in January aimed at closing an offshore cloud rental loophole that gave Chinese companies remote access to restricted U.S. accelerators.
What remains unverified
Moonshot also described a 48-hour autonomous design run in which K3 created a simulated inference chip for a nano model based on its own architecture. Using open-source EDA tools and the Nangate 45nm library, the design closed timing at 100 MHz within 4 square millimeters, included 1.46 million standard cells and an INT4 MAC array, and sustained more than 8,700 tokens per second of simulated decode, according to Moonshot.
For now, K3’s published numbers are Moonshot’s claims or figures drawn from API access. They cannot be checked against local runs until the weights are released. Anthropic accused Moonshot in February of using 3.4 million Claude exchanges to train models through distillation, a claim Moonshot’s K3 benchmarks now place back in the spotlight because the model scores close to systems named in that complaint.
This story draws on original reporting from Tom's Hardware.
