Kimi K2 Takes Open Models to a Trillion Parameters

On July 11, 2025, Moonshot AI published the Kimi K2 Instruct model card and weights. The original source supports the documentary core of the event; total parameter count describes architecture; it is not active memory, speed or quality.

The launch matters less for the number itself than for what it enables: companies and developers can run, adapt, and study a large model without relying entirely on someone else’s API. Kimi K2 is also presented as a model designed for agents—systems that do more than generate text, using tools, writing code, and chaining steps to complete a task.

One trillion parameters, but not all at once

Kimi K2 uses a MoE architecture, short for mixture of experts. Instead of activating the entire network for each token, the system routes processing to a small subset of experts.

Moonshot says that, out of its one trillion parameters, Kimi K2 activates around 32 billion for each token processed. That is a decisive difference: it makes it possible to combine substantial potential capacity without incurring the compute cost of using a dense one-trillion-parameter model for every response. Document supporting the figure.

The idea itself is not new. DeepSeek popularized the strategy with models such as DeepSeek-V3, which has 671 billion total parameters, and Meta has also adopted it in Llama 4. But Kimi K2 raises the bar for open weights to a scale not previously seen among models available for download. Document supporting the figure.

Moonshot has released a base version intended for further fine-tuning, along with Kimi K2-Instruct, designed to follow instructions and use tools. The model supports contexts of up to 128,000 tokens—a window large enough to work with code repositories, extensive documentation, or long task histories. Document supporting the figure.

The bet is on coding agents

The company has centered its presentation on agentic coding tests. Unlike a conventional coding exam, these evaluations measure whether a model can locate a bug in a real project, modify multiple files, run tests, and propose a working solution.

According to Moonshot’s published table, Kimi K2-Instruct scores 65.8% on SWE-bench Verified, a benchmark based on real GitHub issues. The company places it above GPT-4.1 and below the best results attributed to the latest Claude models on the same test. Document supporting the figure.

The company also places it ahead of GPT-4.1 and Claude Opus 4 on some agentic coding tests, including LiveCodeBench. That distinction matters: benchmarks do not crown a universal winner. Results vary depending on the tools allowed, the runtime, the model that plans the task, and the exact version of each evaluation.

Even so, the gap with closed models has narrowed. Two years ago, downloading an open model capable of competing on complex software fixes would have been unthinkable; now, the limiting factor may be the infrastructure required to serve it rather than access to the weights.

Open weights do not mean a restriction-free license

According to the license published by Moonshot, Kimi K2 is distributed under a modified MIT license. It permits broad use, including many commercial applications, but imposes additional conditions on large-scale services: anyone exceeding 100 million monthly active users or $20 million in monthly revenue must seek a commercial license from Moonshot. Document supporting the figure.

That is why it is important to distinguish between open weights and software that is fully free under a standard license with no usage restrictions. For researchers, small companies, and teams that want to deploy the model internally, the license leaves considerable room to operate. For a large consumer platform, Moonshot retains negotiating leverage.

Moonshot also offers Kimi K2 through an API. According to the prices announced by the company, the cost is $0.60 per million uncached input tokens and $2.50 per million output tokens; cached inputs cost $0.15 per million tokens. That is a competitive price for a model of this scale, although running the weights in-house will still require servers equipped with several high-end GPUs. Document supporting the figure.

China strengthens its position in open models

The announcement reinforces a clear trend: several of the most ambitious advances in models with accessible weights are coming from Chinese labs. DeepSeek, Alibaba’s Qwen, and now Moonshot AI have turned open weights and mixture-of-experts models into a way to compete with closed US systems.

Kimi K2 does not erase the advantage held by companies controlling the most powerful proprietary models, nor does it by itself prove that a team can replace them without incurring costs. But it does change the conversation for companies that need control over their data, custom fine-tuning, or independence from a single provider. The next test will be less eye-catching than the announcement: determining whether its benchmark results can be reproduced on real-world projects and whether the model maintains that performance outside carefully prepared environments.

Turning the headline into a check

The first step is to freeze the system's identity. Moonshot AI published the Kimi K2 Instruct model card and weights. A commercial name may cover different revisions, automatic routes and tools. A test record should preserve date, access mode, configuration, permissions and the full output. Without that snapshot, an improvement or failure observed today cannot rigorously be attributed to the version another person will use tomorrow.

Next, turn how to turn a mixture-of-experts card into practical requirements and tests into cases with acceptance criteria. Build a local sample containing easy, ambiguous, long and deliberately impossible tasks. Record the input, the information the model may consult and what outcome would count as sufficient. A vendor-selected demonstration shows possibility; a test set preserved by the user measures reliability.

Autonomy needs a permission ladder. Reading and proposing are not the same as editing, sending or buying. A safer setup begins with read-only access, requires a preview and reserves execution for explicit approval. It also keeps a log and a rollback path. Judge the model by the errors the surrounding system contains, not by the confidence of its plan.

What the record must preserve

Cost and quality must be measured together. A cheap answer that must be reviewed from scratch can cost more than a slower but verifiable one. Measurement includes waiting, retries, consumption, human oversight and the consequences of failure. total parameter count describes architecture; it is not active memory, speed or quality. That boundary turns the announcement into a testable hypothesis rather than a promise to be believed.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to turn a mixture-of-experts card into practical requirements and tests. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.