|
ResearchAudio.io
Kimi K3 Ships Open, With a Revenue Catch
A 2.5x efficiency claim, a 96-shard download, and fine print.
|
|
2.8T
Parameters
|
1M
Token Context
|
2.5x
Compute Efficiency
|
|
|
Moonshot AI just shipped the largest open weight model ever built: 2.8 trillion parameters, a full technical report, and three pieces of training infrastructure, all released within a single day. Then it wrote a license that stays quiet right up until you are big enough to compete with it. That is the part most of the coverage on this skipped.
|
|
Kimi K3 first appeared on July 16, 2026, through Kimi's own apps and API, with Moonshot promising full weights within a couple of weeks. The company beat its own target. Public weights, the technical report, and the supporting infrastructure landed the evening of July 26, roughly a day ahead of the July 27 date Moonshot had originally set, with the company marking July 27 as "Kimi K3 Open Day."
|
|
The jump in size is real. Kimi K3 carries roughly three times the parameters of Kimi K2.5, Moonshot's previous flagship, crossing into a parameter range no open weight model has reached before. Moonshot says that scale came with a 2.5x gain in scaling efficiency under constrained compute, meaning each unit of compute now produces roughly 2.5 times the capability it did last generation, by the company's own accounting, a figure that has not been independently reproduced yet.
|
|
The timing is not an accident. Alibaba has already teased a 2.4 trillion parameter Qwen 3.8 with open weights of its own, and DeepSeek's V4 is expected to leave preview soon. Independent AI commentator Nathan Lambert described the past week as a turning point in the open weight race out of China, and for once that framing undersells it.
|
Where the Efficiency Claim Actually Comes From
|
|
Two changes anchor it. Kimi Delta Attention (KDA), a linear attention mechanism, is mixed with Gated MLA (multi head latent attention) at a 3 to 1 ratio, handling long context with less computation than standard full attention while the latent-attention layers fill in where linear attention loses precision. Attention Residuals (AttnRes) then carries information across layers at the block level instead of the usual per-layer residual connection, acting as a straightforward substitute for it, according to Moonshot.
|
|
Mixture of Experts models activate a subset of a much larger parameter pool for every token, and Kimi K3 pushes that ratio further than before. Its Stable LatentMoE framework activates 16 out of 896 routed experts per token, a sparsity level Moonshot says needed two extra stability techniques (SiTU-GLU and Quantile Balancing) to keep training from breaking down. The vision side, MoonViT-V2, skips contrastive pretraining entirely and trains straight from next-token prediction, reaching similar practical performance to the more common SigLIP-based approach while, Moonshot says, optimizing more smoothly.
|
|
All of it trains under quantization aware training from the supervised fine-tuning stage onward: weights in MXFP4, activations in MXFP8. That differs from the usual approach of training at full precision and compressing afterward, and Moonshot says it holds onto more quality because the model learns to work within the lower precision rather than getting squeezed into it later. The payoff shows up in the download itself: the full model ships at 1.56 terabytes across 96 weight shards, well under the roughly 5.6 terabytes one analysis estimated a full precision version would need.
|
KDA + AttnRes
3:1 mix, long context
|
Stable LatentMoE
16 of 896 experts active
|
MoonViT-V2
vision via next-token prediction
|
Quant-Aware Training
MXFP4 weights, MXFP8 activations
|
|
Source: Moonshot AI, Kimi K3 technical report (Jul 2026)
|
|
|
Moonshot also open sourced the infrastructure behind the training run. MoonEP handles expert-parallel communication for very large, fine-grained Mixture of Experts models without losing efficiency when experts are unevenly loaded. FlashKDA, a fast kernel implementation of Kimi Delta Attention, runs 1.72 to 2.22 times faster than the flash-linear-attention baseline during prefill on Nvidia H20 chips, and Moonshot built it as a direct swap for that baseline. AgentEnv, developed with KVCache.ai, is a sandbox system for running agent training environments at scale, with fast snapshotting and forking built in.
|
|
|
How It Actually Scores
|
|
The benchmark story is genuinely mixed, which is itself useful information. Moonshot compared Kimi K3 against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5, and GLM-5.2 across roughly 40 internal and third party evaluations. K3 leads clearly on tasks built around browsing, tool calling, and pulling answers out of long documents. On general coding and knowledge heavy benchmarks, Claude Fable 5 and GPT-5.6 Sol still come out ahead more often than not.
|
| Benchmark |
Kimi K3 |
Best of Rest |
Edge |
| BrowseComp |
91.2 |
90.4 |
K3 |
| MCPMark-Verified |
94.5 |
92.9 |
K3 |
| FrontierSWE |
81.2 |
86.6 |
Fable 5 |
| GDPval-AA v2 (Elo) |
1686 |
1747 |
Fable 5 |
|
Source: Moonshot AI technical report, benchmarks as of Jul 23, 2026
|
|
|
The harness matters here more than usual. Several of Kimi K3's coding scores run through Moonshot's own Kimi Code harness, while Claude and GPT-5.6 Sol's numbers in the same table often come through Claude Code or Codex instead, so reading every row head to head would overstate how comparable the figures really are. Treat the table as a rough map of where K3 is strong (agentic search, tool use, long document work) rather than a single ranked leaderboard.
|
|
The 48 hour version: if your workload leans on browsing, tool orchestration, or pulling structured answers out of long documents, K3's numbers say it is worth a real evaluation run this week. If you are choosing mainly for general coding or knowledge heavy reasoning, Fable 5 and GPT-5.6 Sol are still ahead on more of those benchmarks.
|
|
The License, Read in Full
|
|
Moonshot does not call Kimi K3 open source, not once, in its own materials. The company uses "open weight" instead, and after reading the Kimi K3 License, that distinction looks deliberate.
|
|
The license starts out broad: download, copy, modify, merge, distribute, sublicense, or build a product on the weights, all without a separate agreement, whether you're a solo developer or a large company. Two conditions kick in past that. If a licensee and its affiliates run a Model as a Service business, meaning outside users get real control over inputs, parameters, or training data rather than just calling a fixed product, and clear 20 million dollars in combined revenue over any trailing 12 months, that group needs a signed agreement with Moonshot before any commercial use. Separately, any commercial product with more than 100 million monthly active users, or more than 20 million dollars in monthly revenue, has to display "Kimi K3" somewhere in its interface.
|
|
Section 4 carves out the exemptions, and this is the part worth reading closely. Purely internal use, meaning the software never reaches anyone outside the licensee, is exempt from both conditions. So is any use that goes through Moonshot's own products or its certified inference partners. A small team running Kimi K3 for internal tooling, or a product simply calling a certified partner's endpoint, never touches either threshold.
|
|
20 million
Dollars in combined revenue, any 12 months, before Moonshot's Model as a Service clause applies
|
|
| Situation |
Threshold |
What Happens |
| Model as a Service reseller |
20M+ combined revenue, 12 months |
Signed agreement required first |
| Large commercial product |
100M+ users or 20M+ monthly revenue |
Must display "Kimi K3" in UI |
| Internal use or certified partner |
No threshold crossed |
No extra step |
|
Source: Kimi K3 License, Sections 2 to 4
|
|
|
This is a tighter design than Meta's old approach with Llama, which drew its commercial line at 700 million monthly users, a bar few products ever cross. Moonshot's line is measured in revenue, not users, and it targets a much smaller business: a Model as a Service reseller clearing 20 million dollars a year looks nothing like a company with 700 million users. DeepSeek, for comparison, ships its models under a plain MIT license with none of these conditions attached.
|
|
The 48 hour version: before you put Kimi K3 behind a customer facing feature, check which row of that table you land on. Calling a certified partner's endpoint or running it for internal tooling puts you in the bottom row, with no extra step. The 20 million dollar figure is aimed at whoever might resell K3 access at scale, and OpenRouter already lists it through 7 providers at 3 dollars per million input tokens and 15 dollars per million output tokens, the same rate Moonshot charges directly.
|
|
|
|
|
Together AI and Modal both shipped hosted access to Kimi K3 the same day the weights went up, and OpenRouter now lists it through 7 different providers.
|
|
Moonshot recommends supernode setups with 64 or more accelerators for serving K3, a direct result of how much cross chip communication its Mixture of Experts routing needs.
|
|
Kimi K3 always reasons by default: the reasoning_effort field takes low, high, or max, and max is the default. Multi-turn tool calls need the full previous response, including reasoning_content and tool_calls, passed back unchanged, or the model loses track.
|
|
Alibaba has teased a 2.4 trillion parameter Qwen 3.8 with open weights of its own, and DeepSeek's V4 is expected to leave preview soon. This will not be the last model in this size class this year.
|
|
|
|
Here's the part nobody's talking about: that 20 million dollar threshold is not really written for the person reading this newsletter. It's written for the layer above you, the cloud and inference resellers who would otherwise host K3 and undercut Moonshot's own pricing on day one. Read the license anyway, not because the headline number applies to you, but because Section 4's exemptions are doing more of the real work than the number everyone is quoting.
|
|
|
|
Moonshot never calls this open source, not once. If a revenue tiered license is how frontier labs stay open without subsidizing resellers, does "open weight" need its own tiers, the way cloud pricing already does? And would a flat MIT license, like DeepSeek's, survive if its models reached this same scale?
|
|
The full 40-benchmark table and the vLLM and SGLang deployment configs are in the paid archive.
|
|
The interesting split in open models right now is not open versus closed anymore. It's who the openness is actually for. Next: what happens when Alibaba's 2.4 trillion parameter Qwen 3.8 lands, and whether it ships with a license like this one or a plain MIT file.
|
|
ResearchAudio.io
Sources: Moonshot's Kimi K3 announcement, the Kimi K3 License, Nathan Lambert's analysis at Interconnects
|
