Kimi K3 announced: 2.8T MoE, 1M context, weights due July 27 under Modified MIT
Contents
Moonshot AI announced its flagship model Kimi K3 on July 16, 2026.
It is a 2.8-trillion-parameter MoE, and the announcement calls it “the largest open-weight model ever.”
The weights are not out yet, though.
The stated plan is to release them by July 27 under a Modified MIT license (commercial use allowed); for now the only way to use K3 is the API or kimi.com.
I have covered the previous Kimi releases on this blog.
K2.6 from April is in the Qwen3.6-Max-Preview comparison, and K2.7 Code from June is in the Chinese agent models post.
K3 is the release that follows those two.
What was announced
| Item | Kimi K3 |
|---|---|
| Total parameters | 2.8T (MoE, 16 of 896 experts active) |
| Context length | Native 1M tokens (4x the 256K of K2.6) |
| Input/output | Text + image input, text output |
| License | Modified MIT (weights promised by July 27) |
| API | Model ID kimi-k3, $3.00/M input, $15.00/M output |
The MoE design is called “Stable LatentMoE”: 896 experts (specialist subnetworks), of which 16 run at inference time.
Only a small slice of the 2.8T total is active per token — the same direction as the K2 family, at roughly 3x the size.
New architecture pieces
The headline component is Kimi Delta Attention (KDA).
It is a hybrid of regular attention and linear attention (an approximation whose cost grows linearly with context length), and Moonshot claims up to 6.3x faster decoding at 1M-token context.
The native 1M context depends on this mechanism.
The other one is Attention Residuals (AttnRes).
This is the method Moonshot put on arXiv in March that replaces Transformer residual connections with depth-wise attention — I wrote an explainer about it at the time.
Four months after the paper, it is now in the flagship.
In K3 they claim about a 25% training-efficiency gain for under 2% extra cost.
Beyond that, K3 uses Gated MLA, Sigmoid Tanh Unit (SiTU), Quantile Balancing for router optimization, and Per-Head Muon.
Training ran with MXFP4 weights and MXFP8 activations, and scaling efficiency is stated as roughly 2.5x over K2.
Details like the training token count are pending the technical report, which is due alongside the weight release.
Benchmarks and positioning
Excerpted from the official announcement and press coverage (source: MarkTechPost).
| Benchmark | K3 | Claude Fable 5 | GPT-5.6 Sol | Opus 4.8 | GLM-5.2 |
|---|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 46.2 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 63.7 |
| Terminal-Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 67.3 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 13.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | — |
Moonshot itself states in the announcement that K3 “still trails Claude Fable 5 and GPT-5.6 Sol overall.”
At the same time it claims first place on SWE Marathon (long-horizon coding) and Program Bench, and SOTA on BrowseComp at 91.2.
The loudest post-launch story was Arena.AI’s Frontend Code Arena, where K3 took first place at 1679 points.
On Artificial Analysis, the Intelligence Index is 57, around 4th place out of 189 models.
Generation speed is below average at 62 tokens/sec, and the output is rated as very verbose.
Pricing and availability
The API is at https://api.moonshot.ai/v1 with model ID kimi-k3.
Pricing is $3.00/M input on cache miss, $0.30/M on cache hit, and $15.00/M output.
That is the most expensive a Chinese lab has priced so far — Simon Willison calls it “Claude Sonnet-class pricing.”
Moonshot claims a cache-hit rate above 90% on coding workloads; if that holds, the effective input price drops well below the sticker price.
Beyond the API there is the kimi.com chat, Kimi Code for terminal/IDE, the Kimi Work desktop app, mobile (iOS/Android/HarmonyOS), and Kimi Enterprise.
The reasoning-token overhead
On Willison’s standard test (the pelican-on-a-bicycle SVG), a 95-token input burned 13,241 reasoning tokens, about 25 cents for one generation.
That lines up with the “very verbose” rating from Artificial Analysis.
Whether reasoning volume can be dialed down manually is unknown right now, and commenters have also questioned an estimated 85-token hidden system prompt on transparency grounds.
The per-generation cost is the $15/M output price multiplied by this reasoning volume.
The weight release and the size debate
The Hacker News thread (1,730 points, 1,000+ comments) split over whether calling a 2.8T model “open weights” means much in practice.
Even at FP8 the weights alone come to roughly 2.8TB, so effectively no one can run it locally.
The argument goes that once the weights land, most users are still on the API either way.
By contrast, June’s K2.7-Code (1T total / 32B active MoE) is already on Hugging Face under Modified MIT, with over 780,000 downloads.
The moonshotai Hugging Face org currently stops at K2.7-Code; there is no K3 checkpoint yet.
Note that the original K2-line API was retired on May 25, 2026; in release order the line runs K2 (July 2025) → K2.5 → K2.6 → K2.7-Code → K3.
The general-benchmark numbers like AIME and MMLU, the training token count, and whether smaller distilled variants exist should all be filled in by the weights and the technical report due by July 27.