Moonshot AI has announced its flagship model "Kimi K3," with a total of 2.8 trillion parameters. It supports a 1-million-token context and visual input, and runs this enormous model sparsely via a Mixture of Experts (MoE) architecture that selects 16 out of 896 experts. The API is already available, but the model weights and technical report are scheduled to be released by July 27, 2026. What matters in practice is the combination of a design that decouples total parameter count from the computation used per inference, API pricing lower than closed models, and deployment requirements that recommend 64 or more accelerators.
Even at 2.8 Trillion, Not Everything Runs Every Time
Kimi K3's 2.8 trillion figure refers to the total number of parameters the model holds. It does not mean all 2.8 trillion are computed for every token processed. K3 selects 16 out of 896 experts, routing computation to the paths that best fit the input. The selection ratio is about 1.79%, but multiplying this by 2.8 trillion does not yield the number of active parameters, since attention layers, embeddings, and shared components exist separately.
Comparing K3 to the original Kimi K2 makes it easier to see where Moonshot AI expanded things.
| Item | Kimi K2 | Kimi K3 |
|---|---|---|
| Total parameters | 1 trillion | 2.8 trillion |
| Number of experts | 384 | 896 |
| Experts selected per token | 8 | 16 |
| Active parameters | 32 billion | Not disclosed |
| Context length | 128,000 | 1,048,576 |
The total scale is 2.8 times that of K2, and the pool of expert candidates has more than doubled. Even so, K3's active parameter count has not been disclosed. Model capacity, inference-time computation, and measured speed cannot all be described by a single "2.8 trillion" figure. It remains to be seen whether the July 27 technical report will reveal the active parameter count.
Moonshot AI states that it trained the model with quantization in mind from the supervised fine-tuning stage onward, using MXFP4 for weights and MXFP8 for activations. A simple conversion of 2.8 trillion weights into 4-bit values alone comes out to roughly 1.4TB. In actual operation, quantization format metadata and activations must also be retained, and KV cache and runtime memory are needed as well. While sparse MoE can reduce the computation performed per operation, it does not eliminate the problem of storing the entire model and delivering data to experts at high speed.
KDA and AttnRes: Untangling Long Context and Depth Separately
At the core of K3 are Kimi Delta Attention (KDA), which handles sequence length, and Attention Residuals (AttnRes), which addresses the network's depth dimension. KDA combines ordinary attention with linear attention that uses finite states. In Moonshot AI's prior research, Kimi Linear—with 48 billion total and 3 billion active parameters—reduced KV cache by up to 75% compared to a full MLA configuration and achieved up to 6x faster decoding at 1 million tokens.
However, that 75% and 6x figure were obtained with Kimi Linear. The same multipliers cannot be directly applied to the 2.8-trillion-parameter K3. Moonshot AI has prepared a vLLM implementation of prefill caching compatible with KDA, which it plans to release alongside the model weights. It explains that this caching reuses long shared segments to keep down token pricing on the official API. It claims cache hit rates exceed 90% under coding workloads, though results will vary depending on how users structure their prompts.
AttnRes, rather than having every Transformer layer uniformly add the output of previous layers, re-selects necessary past representations via depth-wise attention. Since retaining all layers increases memory load, K3 uses a practical configuration that groups layers into blocks. In prior experiments, Block AttnRes reached a loss equivalent to training standard residual connections with 1.25x the computation. This too is a result from a smaller precursor model rather than K3 itself. The claim that K3 achieves "roughly 2.5x overall scaling efficiency compared to K2" needs to await the release of the technical report, including its measurement formula.
As an example of long-horizon work, Moonshot AI had K3 design a chip for small models. In a single 48-hour run, K3 reportedly completed a design using open-source EDA and the Nangate 45nm library, achieving 4mm², 100MHz, and a simulated throughput exceeding 8,700 tokens per second. The design incorporates 1.46 million standard cells, 0.277MB of SRAM, and INT4 arithmetic units. This is not a manufactured semiconductor but a proof of concept spanning logic design through verification.
An API 40-50% Cheaper, With Benchmark Harness Differences
Kimi K3's API pricing is $3 per million tokens for input not in cache, and $15 for output. The same rate applies across the entire 1-million-token context, and cached input drops as low as $0.30.
| Model | Input | Output |
|---|---|---|
| Kimi K3 | $3 | $15 |
| GPT-5.6 Sol | $5 | $30 |
| Claude Fable 5 | $10 | $50 |
All figures are published prices per million tokens. K3 is 40% cheaper on input and 50% cheaper on output than GPT-5.6 Sol. Compared to Claude Fable 5, the gap is 70% for both input and output. Furthermore, once input on GPT-5.6 Sol exceeds 272,000 tokens, the input rate for the entire request doubles and the output rate rises 1.5x. For use cases that pass long codebases or document sets every time, K3's flat pricing and automatic caching make budgeting easier.
Performance cannot be compared as straightforwardly as the price table. According to Moonshot AI's published figures, K3 scored 88.3 on Terminal Bench 2.1, versus 88.8 for GPT-5.6 Sol. On BrowseComp, K3 scored 91.2, surpassing GPT-5.6 Sol's 90.4 and Claude Fable 5's 88.0. Meanwhile, on the Elo score for GDPval-AA v2, which measures professional tasks, K3 scored 1,668, GPT-5.6 Sol scored 1,748, and Fable 5 scored 1,760.
These numbers were not measured side-by-side in the same execution environment. Different agent execution frameworks—KimiCode, Claude Code, and Codex—are mixed in, and some values were taken from existing leaderboards. Fable 5 may switch to a different model, and GPT-5.6 Sol may trigger safety mechanisms (cyberguard). Moonshot AI itself acknowledges that K3 falls short of both models in overall performance and that there are clear differences in usage experience. What K3 demonstrates is not an across-the-board top position, but a place among the top tier on several long-horizon agentic tasks, achieved at a lower price.
What "Open" Really Means Will Be Tested on July 27
Kimi K3 is currently available on Kimi.com, Kimi Work, Kimi Code, and the official API. The complete model weights and technical report are scheduled to be released by July 27, and the distribution license has not yet been announced. Moonshot AI refers to the model as "open source" even including the future weight release, but for now API availability comes first. The stage where third parties can obtain the model and reproduce its total parameter count or inference performance has not yet been reached.
The Open Source Initiative's definition requires freedom to use, study, modify, and share. It further requires providing sufficient training data information, along with the code used for training and running the model, and the model's parameters in a form suitable for modification. Even once weights are released, that alone does not guarantee satisfying the OSI's Open Source AI Definition. Whether K3's released materials go far enough—including licensing and training information—will need to be verified.
The barrier to self-hosting is also high. Moonshot AI recommends a supernode equipped with 64 or more accelerators for deploying K3. Given the simple estimate of roughly 1.4TB for 4-bit values alone, weight availability and whether it can be run economically are separate issues. The affordable official API is also the result of building KDA and quantization into the model, and optimizing a disaggregated inference infrastructure and caching on the service side.
Operational quirks have also already been documented. K3 always has thinking mode enabled, and currently the only supported reasoning_effort value is max. Across multiple turns, quality becomes unstable unless the entire response—including past reasoning—is fed back into the next request, and switching to a different model midway through a conversation is not recommended. Because the model may make unexpected judgments on the user's behalf when given ambiguous instructions, official documentation encourages specifying the scope of behavior clearly via system prompts or AGENTS.md.
July 27 will be a day to check the scope of the release, rather than the headline figure of 2.8 trillion total parameters. First, the active parameter count and license are needed. If training information and evaluation conditions reproducible by third parties are also provided, K3 will progress from a low-priced API to a verifiable open model. That scope of disclosure will affect developers' choices far longer than the banner of being one of the largest models.