The drop
On July 16, 2026 Moonshot released Kimi K3, and the headline number is the size: 2.8 trillion parameters, which Moonshot says makes it “the first open model to reach 2.8 trillion parameters.” It is a Mixture-of-Experts architecture - a “Stable LatentMoE” framework that activates 16 of 896 experts per token - built on two mechanisms Moonshot is betting the house on: Kimi Delta Attention (a hybrid linear attention scheme meant to scale cheaply) and Attention Residuals, which retrieve representations across depth instead of the usual residual connections. Training was quantization-aware from the SFT stage onward, with MXFP4 weights and MXFP8 activations.
The practical specs are the ones that matter for buyers: a native 1,000,000-token context window, native vision (text, images, and video in the same model), an always-on thinking mode (locked to “max” effort at launch, with low and high modes promised later), full function and tool calling with MCP, and automatic prompt caching that Moonshot says hits above 90% on coding workloads.
Full open weights are promised by July 27, 2026. As of launch day the HuggingFace repo is not populated yet, so “open” is a commitment, not a download. See the model card: tokenstead.ai/models/kimi-k3.
Where it actually lands
Two independent leaderboards, not Moonshot’s own numbers:
-
Artificial Analysis webdev arena - #1. Kimi K3 tops the webdev code leaderboard at a score of 1679, ahead of Claude Fable 5 (1631) and GPT-5.6 Sol (1618). It carries a “Preliminary” tag on roughly 1,757 votes, so the rating can still move - but it is genuinely leading the board today.
-
Artificial Analysis Intelligence Index - #4 of 189, score 57. That puts it in the neighborhood of Opus 4.8 and GPT-5.5, behind Fable 5 and GPT-5.6 Sol. Artificial Analysis also flags it as slower than average (62 tokens/sec, #90 of 189) and “very verbose.”
Moonshot’s own benchmarks (all at reasoning effort “max”) are strong but self-reported: GPQA-Diamond 93.5, HLE 43.5, Terminal Bench 2.1 88.3, FrontierSWE 81.2, DeepSWE 67.5, BrowseComp 91.2, MMMU-Pro 81.6. Head to head against the proprietaries Moonshot publishes, K3 beats Opus 4.8, GPT-5.5, and GLM-5.2 across most benches, and trails Fable 5 and GPT-5.6 Sol overall. It does not beat the top two - but it is in the conversation, which is the story.
The pricing signal
Here is the part that tells you what Moonshot thinks K3 is:
-
$3.00 per 1M input (cache miss)
-
$0.30 per 1M input (cache hit)
- $15.00 per 1M output
That is roughly 5x what Kimi K2.7 cost. Moonshot’s entire brand identity was “the cheap one” - the lab that put frontier-ish quality at budget prices. With K3 that deal is over. They are charging like it competes with Opus, because that is the tier they are claiming. At this price, “pretty good for the cost” stops being the pitch; the model has to hold the frontier on merit, or the price is a joke. The webdev arena result is the evidence they are pointing at.
Self-hosting reality
The “open” in open-weight needs an asterisk for almost everyone. Recommended inference is a supernode with 64 or more accelerators. Even once the weights drop on July 27, that is not something you put on a workstation or a Mac the way you can with a 30B dense model. For practical purposes K3 is cloud-only - you call it through Moonshot’s API or a provider, you do not run it yourself. The openness matters for researchers and for the labs that will fine-tune on it; it does not change the self-hosting picture for an individual developer.
The takeaway
K3 is a serious model at a serious price, and the independent evidence (webdev #1, Artificial Analysis #4) backs the seriousness. The two open questions are whether the weights actually land on July 27, and whether the frontier-level pricing holds up once Fable 5 and GPT-5.6 Sol respond. For now, if you are doing web development and paying per token, it is worth pointing your harness at the kimi-k3 model card and running your own eval.