(Source: PHOTOCREO Michal Bednarek/Shutterstock)
Moonshot AI, a Beijing startup backed by Alibaba, has released a new open weight large language model. Called Kimi K2, the model has one trillion parameters, making it one of the largest open weight models available. The system activates 32 billion parameters at a time and is offered in both a foundation model and instruction-tuned version.
These are the model descriptions from Moonshot:
Kimi-K2-Base: The foundation model, a strong start for researchers and builders who want full control for fine-tuning and custom solutions
Kimi-K2-Instruct: The post-trained model best for drop-in, general-purpose chat and agentic experiences. It is a reflex-grade model without long thinking
Kimi K2 is designed for code generation and agentic problem-solving, meaning it can decide which functions to call, evaluate the results, and adjust its plan without fresh prompts. It uses a mixture-of-experts (MoE) architecture, a neural network design that breaks a large model into many smaller “expert” subnetworks, each trained to handle a particular segment of the data or task. A gating layer decides which experts should process each input, so only a fraction of the total parameters runs at any moment, trimming the training and inference compute needed to run the model.
This efficiency translates to big savings: Kimi K2’s pay-as-you-go pricing is about $0.15 per million input tokens and $2.50 per million output tokens, sitting well below most frontier models. OpenAI’s GPT-4.1, for example, lists $2.00 per million input tokens and $8.00 for output, while Anthropic’s Claude Opus 4 comes in at $15 and $75. These savings can translate into thousands of dollars for users who process millions of tokens per day, giving Kimi a cost profile more typical of mid-tier models despite its trillion-parameter scale.
But price is only part of the appeal: Initial tests show that Kimi K2 matches or beats the performance of leading proprietary models on key coding and math benchmarks. On SWE-Bench Verified, a human-validated subset of the SWE-Bench software engineering benchmark, Kimi K2 scores 65.8% vs. GPT-4.1’s 54.6%. This benchmark contains 500 real GitHub bug-fix tasks that an AI model must solve by generating a patch that passes all unit tests. Kimi K2’s 65.8% score means it automatically repaired almost two-thirds of those issues. On LiveCodeBench, an end-to-end coding benchmark, Kimi K2 reaches 53.7% vs. GPT-4.1’s 44.7%. In math reasoning (MATH-500), it hits 97.4%, slightly above GPT-4.1’s 92.4%.
This graphic provides a snapshot of Kimi K2’s benchmark performance. (Source: Moonshot AI)
Moonshot AI credits its new MuonClip optimizer with keeping Kimi K2’s trillion-parameter training run completely stable. Large transformer models often suffer from “logit explosions” in the attention layers, which can crash a job or force costly restarts. MuonClip tackles the problem at its root: After every update, it rescales the query and key weight matrices, keeping the raw attention values in a safe numerical range and preventing the runaway growth that derails many large models during training.
“Our experiments show that MuonClip effectively prevents logit explosions while maintaining downstream task performance. In practice, Kimi K2 was pre-trained on 15.5T tokens using MuonClip with zero training spike, demonstrating MuonClip as a robust solution for stable, large-scale LLM training,” the company wrote. By preventing training crashes and cutting compute waste, MuonClip lets researchers share checkpoints and compare training methods on a level field, which could significantly lower the cost of large-scale experiments.
The launch of Kimi K2 comes as Beijing urges domestic firms to close the gap with U.S. rivals while Washington tightens export rules on advanced chips. By releasing the weights of this trillion-parameter model, Moonshot and its peers like Deepseek aim to build global developer trust while mitigating local hardware constraints. U.S. companies have moved in the opposite direction: OpenAI and Anthropic cite safety concerns for holding back their newest weights, leaving Meta’s Llama line as the main Western open model at scale.
Founded in 2023 by Tsinghua University graduate Yang Zhilin, Moonshot first gained attention with a chatbot that handled extremely long inputs at the time. Kimi K2 extends that work with stronger coding skills and a context window of up to 128,000 tokens. Moonshot is offering access to Kimi K2 through its web and mobile products and an API.
Moonshot noted the model does have some limitations: “In our internal tests, we’ve identified some limitations in current Kimi K2 models. When dealing with hard reasoning tasks or unclear tool definitions, the model may generate excessive tokens, sometimes leading to truncated outputs or incomplete tool calls. Additionally, performance may decline on certain tasks if tool use is enabled. When building complete software projects, one-shot prompting yields performance degradation compared to using K2 under an agentic framework. We are working to address these issues in future releases and looking forward to more feedback,” the company wrote. Read more about the technical aspects of Kimi K2 and see demos of its capabilities at this link. Try the model for free here.
Open weight models at the trillion-parameter scale mark a turning point from private experimentation to shared scientific infrastructure. This aligns with the objectives of the Trillion Parameter Consortium, an organization that aims to build an open community, identify and incubate promising ideas to facilitate their growth, and create a global network of expertise and resources.
Taking place from July 29-31, TPC25 is the Trillion Parameter Consortium’s 2025 all-hands conference and exhibition, convening AI leaders from industry, academia, and national laboratories along with the vendor community, funding agencies, and VCs, to develop best practices for utilizing AI for scientific discovery and engineering at scale. Learn more about the conference and how to participate by visiting TPC25.org.