On July 16, 2026, Moonshot AI released Kimi K3 to its API and consumer products. The model has 2.8 trillion total parameters, making it one of the largest open-weight AI models ever released. That headline is real. Most of the coverage stops there. The more useful questions are what the architecture actually does, what the benchmark table is hiding, and why a Chinese lab is charging Anthropic Sonnet prices.
What Kimi K3's architecture actually does
Kimi K3 has 2.8 trillion total parameters with aggressive mixture-of-experts sparsity - 16 of 896 experts active per token under a Stable LatentMoE framework - a 1-million-token context window, and native multimodality covering images and video. The 2.8T figure is a total-parameter count. What runs per token is far smaller. Stable LatentMoE activates only 16 of 896 experts for each token, meaning just 1.8% of experts are used at a time. This is what makes a model at this scale remotely serviceable, but it introduces hard routing and memory problems that less sparse architectures avoid.
Two architectural pieces are new and worth understanding separately. Kimi Delta Attention is credited with up to 6.3x faster decoding in million-token contexts, and Attention Residuals are credited with roughly 25% higher training efficiency at under 2% additional cost - both vendor-stated figures. KDA targets the long-context serving problem: Kimi Delta Attention is the efficient-attention foundation, which Moonshot positions as a way to scale long-sequence processing without carrying the full cost profile of ordinary attention. Attention Residuals work along the depth axis: AttnRes selectively retrieves representations across depth rather than accumulating them uniformly.
Together with improvements in training methodology and data recipes, these structural advances give Kimi K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively.
One practical note for teams who want to self-host once the weights land: Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators for inference efficiency.
Quantization-aware training from the SFT stage uses MXFP4 weights with MXFP8 activations for broad hardware compatibility. This is a data-center decision, not a lab server.
What the benchmark table actually says
Artificial Analysis gives K3 a score of 57 on its Intelligence Index and ranks it fourth among 189 models. It sits behind Claude Fable 5 and two GPT-5.6 Sol reasoning settings, then ahead of Claude Opus 4.8, GPT-5.5 at xhigh, Claude Sonnet 5, and GLM-5.2.
That's genuinely strong. But the benchmark table Moonshot published mixes three different evaluation harnesses - KimiCode, Claude Code, and Codex - depending on which model is being tested. Moonshot's detailed benchmark table mixes Kimi Code, Claude Code and Codex harnesses; several K3 results were not yet visible on the referenced public leaderboards; and the promised model weights are not available today. Running K3 under its own harness and competitors under theirs is the most common way benchmark tables overstate a model's lead. The independent Artificial Analysis score, which runs all models under the same methodology, is the number to anchor on.
Where K3 looks genuinely competitive: K3 leads or matches frontier models on Terminal Bench 2.1 (88.3) and SWE Marathon (42.0).
It trails Fable 5 on FrontierSWE and DeepSWE but beats Opus 4.8 and GPT-5.5 on most coding suites.
The verbosity problem is real and directly affects agent cost. When evaluating the Intelligence Index, K3 generated 130M tokens, which is very verbose in comparison to the average of 63M. Right now, reasoning effort currently supports only the max level (default); more levels are coming soon. K3 always reasons at full depth, which generates long internal chains regardless of task complexity. That matters a lot at $15 per million output tokens.
Beagle in actionengineering channel, 10:02amThe pricing story nobody is telling
The new model is notable for the pricing: $3/million input tokens and $15/million output tokens, putting it at the same level as Anthropic's Claude Sonnet series and making it the most expensive model released by a Chinese AI lab to date.
Compare that to where Moonshot was six months ago. This is a significant increase on their earlier models such as Kimi K2.6 at $0.95/$4.
K3 costs about 3.2 times as much for fresh input and 3.75 times as much for output as K2.6 or K2.7 Code.
The useful number is not the list price, though. Leveraging Mooncake disaggregated inference infrastructure, Kimi reports over 90 percent cache hit rates for development workloads. The cached-input rate is $0.30 per million - one tenth of the uncached rate. For coding agents running against stable repository context, the effective blended cost is a fraction of the headline. The list price still hides the number that matters for agents. K3 always reasons, and the only launch setting is max. A $15 output rate becomes expensive when a task generates a long reasoning trace, retries failed tool calls, or drags its history through many turns.
| Model | Input ($/M) | Output ($/M) | Cache hit ($/M) | AA Index |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 | 57 |
| Claude Fable 5 | $10.00 | $50.00 | ~$1.00 | ~60 |
| Claude Sonnet 5 | $3.00 | $15.00 | $0.30 | ~45 |
| GLM-5.2 (MIT) | ~$1.40 | ~$5.60 | - | ~50 |
| Kimi K2.6 | $0.95 | $4.00 | - | ~32 |
Prices from Moonshot pricing page and Artificial Analysis; AA Index scores are rounded and model-best configurations.
The practical read: K3 undercuts Fable 5 significantly but matches Sonnet 5. For teams currently on Sonnet 5, K3 is a lateral move on price with a capability gain - worth testing. For teams already on the cheapest Chinese models like K2.7 Code, K3 is a 3-4x cost increase for 15 Artificial Analysis points. That math only works for tasks where general intelligence and long-context reasoning are actually the bottleneck.
The open-weight caveat that matters
Moonshot calls K3 the first open 3T-class model. As of July 17, the Moonshot AI Hugging Face organization listed K2-series models but no Kimi K3 checkpoint. Artificial Analysis therefore classified the currently accessible K3 service as proprietary and explicitly said the weights were unavailable.
Moonshot says it will release K3's weights on July 27, meaning developers cannot independently inspect, modify or run the model themselves yet. There is also no published license, model card, or technical report as of launch day. Open source also has several layers. Public weights permit local inference and research; open training code, reproducible data, a permissive license, and complete technical documentation are separate questions.
This is not a reason to ignore K3. It is a reason to treat it as a hosted API today and wait ten days before making any architectural commitment. If you are building infrastructure that depends on self-hosted weights, the July 27 date is when to re-evaluate, not when to commit.
Kimi K3 open-weight model: common questions
What is Kimi K3 and who made it?
Kimi K3 is a language model from MoonshotAI, released in July 2026, with multimodal input, a 1.0M-token context window, and pricing from $3.00/M input and $15.00/M output. It is Moonshot AI's flagship open Mixture-of-Experts model for long-horizon coding, knowledge work, and reasoning. Moonshot AI is a Beijing-based lab.
How does Kimi K3 rank against Claude and GPT-5?
Artificial Analysis Intelligence Index v4.1 scores K3 Max at 57.1, GPT-5.6 Sol Max at 58.9 and Fable 5 Max with Opus 4.8 fallback at 59.9. K3 is effectively third by model family, two points behind Sol and three behind Fable 5, on an independent benchmark that does not use vendor-supplied harnesses.
When does Kimi K3 release its open weights?
It's currently available via their website and API, but an open-weight release is promised "by July 27, 2026." The license terms, model card, and technical report have not been published as of July 17. Treat the API as a hosted, proprietary service until the checkpoint appears.
How much does Kimi K3 cost per million tokens?
The official rates are $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens. For coding agents with high cache reuse, the effective blended input cost is much closer to $0.30. Output token cost is unchanged and becomes the dominant expense for reasoning-heavy tasks.
Can I run Kimi K3 locally?
Not yet. Weights have not shipped as of July 17. When they do, Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators for inference efficiency. Self-hosting at this scale requires data-center infrastructure; no consumer or workstation path exists for the full model.