Kimi K3: the largest open-weight model ever released has 2.8T parameters

On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model the company describes as the first open model in the 3-trillion class. It is the largest open-weight model ever released, and it arrived roughly a month after the U.S. government pulled Anthropic's Fable and Mythos models from circulation over security concerns. The result is uncomfortable for anyone following the race closely: the most capable frontier model anyone can freely obtain today is Chinese.

What Moonshot actually shipped

K3 uses a mixture-of-experts architecture: of the model's 896 experts, only 16 are activated per token, inside a framework Moonshot calls Stable LatentMoE and on top of its own attention mechanism, Kimi Delta Attention. The context window is 1,048,576 tokens, and the company claims roughly 2.5 times the scaling efficiency of the previous generation.

The generational difference is worth spelling out, because it is easy to conflate. Kimi K2, from July 2025, had 1 trillion total parameters, 32 billion active per token and a 128,000-token context window. K3 is a different tier of scale, not a point release.

One detail that tends to get filled in wrongly: Moonshot has not published an official active-parameter count per inference. Third-party estimates land between 40 and 100 billion, and the naive arithmetic (2.8T times 16/896) gives something close to 50 billion. Any specific figure you read elsewhere is an estimate, not a company number.

Where K3 wins, and where it does not

In the numbers Moonshot published itself, K3 leads two benchmarks that reflect real work: SWE Marathon, at 42.0 against 40.0 for Claude Opus 4.8 and 39.0 for GPT-5.6 Sol, and BrowseComp, at 91.2 against 90.4 for Sol and 88.0 for Claude Fable 5. These are self-reported results, so they deserve the usual scepticism until independently validated.

On Terminal-Bench 2.1, the company's own table shows 88.3 for K3 against 88.8 for GPT-5.6 Sol. That is a statistical tie, not a lead, and Vals' independent evaluation placed K3 second on that same test.

The most useful independent measurement comes from Artificial Analysis, which scored K3 at 57 on the Intelligence Index v4.1. That makes it the most capable open-weight model available, behind Claude Fable 5 and GPT-5.6 Sol, and level with Opus 4.8 and GPT-5.5. The honest reading is the one Moonshot wrote into its own announcement: overall performance still trails the most powerful proprietary models.

"Open-weight" does not mean you will run this

This is where the conversation usually goes off the rails. Open weights sounds like "install it on your own infrastructure and solve the privacy problem." At this scale, that is not what happens.

At 4 bits, in the MXFP4 format the model was actually trained for, 2.8 trillion parameters take up around 1.4 TB in weights alone, before any activations or context cache. And Moonshot's own recommendation for serving the model is configurations of 64 accelerators or more, not because of memory but because of the interconnect bandwidth needed to route every token across experts. Reuters quotes Ryan Fedasiuk of the American Enterprise Institute saying that running this locally would require hundreds of thousands of dollars in equipment.

Add the real state of the openness: as of July 26 the weights were still not public (Moonshot promised release by the 27th), the licence has not been published, and there is no training data or training code. It is open-weight, not open-source, and the difference matters. For a company, what openness concretely delivers is auditability, no single-vendor dependency on the artifact, and freedom to choose who you rent inference from. It is not running it in the server room.

The price breaks the cheap-Chinese-model assumption

The official table is US$0.30 per million input tokens on a cache hit, US$3.00 on a miss, and US$15.00 per million output tokens, with no long-context surcharge anywhere across the 1 million tokens.

On average cost per task as measured by Artificial Analysis, K3 comes in around US$0.94, against US$1.04 for GPT-5.6 Sol and US$2.75 for Fable 5. But it sits well above GLM-5.2 (US$0.47) and Grok 4.5 (US$0.31). In other words: K3 is not the cheap Chinese model of the previous cycle. It is priced in mid-tier proprietary territory, which turns the decision into one about capability rather than savings.

The number that is not in the announcement

On Artificial Analysis' AA-Omniscience evaluation, K3's hallucination rate rose to 51%, against 39% for the previous generation. More capable and, at the same time, more willing to fabricate. The model is also notably verbose: it consumed 130 million output tokens to complete the Intelligence Index suite, roughly twice the median.

Moonshot documented two behavioural limitations in its own announcement, which is rare and deserves credit. The first is sensitivity to reasoning history: if the application discards prior reasoning, or if you switch models mid-session, quality can become highly unstable. The second is excessive proactiveness: faced with ambiguous intent, the model may make unexpected decisions on the user's behalf.

For anyone about to put a model in front of customers, those three facts are worth more than any benchmark. Capability without a verification layer is not a gain, it is a liability. A model that errs more confidently needs more checking, not less.

What to do with this in practice

The actionable path is the API, and it is compatible with the OpenAI SDK: the official documentation uses from openai import OpenAI pointed at https://api.moonshot.ai/v1 with the id kimi-k3. In practice, testing the model against something already running on OpenAI means changing two lines of configuration, which drops the cost of experimenting to almost nothing.

Three concrete recommendations. Test on your own data and in your own language, because none of these benchmarks measure that. Do not put the model in customer-facing contact without a factual verification step in between, given the hallucination rate. And watch July 27: the weights release and, above all, which licence comes with it will determine whether this is a real architectural option or just one more API.

One last data point for a sense of the moment. Demand for K3 was heavy enough that Moonshot suspended new subscriptions, and on launch day its Hong Kong-listed Chinese competitors fell hard, with Zhipu down 27.7% and MiniMax down 16.5%. The company is closing a round valuing it at more than US$30 billion, with US$300 million in annual recurring revenue as of June and a Hong Kong IPO expected within six months.

Sources

• Kimi K3: Open Frontier Intelligence, Moonshot AI official blog
• Kimi K3 Quickstart, official API documentation
• China's Moonshot unveils world's largest open AI model, Reuters
• Kimi K3: Intelligence, Performance & Price Analysis, Artificial Analysis
• Kimi K3: benchmarks, pricing, hardware requirements and self-hosting, Northflank
• Moonshot plans Hong Kong listing at a $30 billion valuation, The Next Web