On July 16, 2026, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter open Mixture-of-Experts model with native vision and a 1-million-token context window. Moonshot calls it the world's first open 3T-class model, and the claim is hard to argue with: for nine of the past twelve months, Kimi releases have set the ceiling on open-model size. K3 pushes that ceiling higher and, for the first time, lands the company's benchmark numbers within striking distance of the closed frontier.
The weights don't drop until July 27, under a Modified MIT license. But the model is already live on Kimi.com, Kimi Code, and the API — so the reception has been immediate, and loud.
A 2.8T model that only fires 16 experts
The headline number is 2.8 trillion parameters, but K3 never runs all of them at once. It uses Stable LatentMoE, a sparse routing scheme that activates just 16 of 896 experts per token. That's the trick behind every large MoE: enormous capacity on disk, a fraction of the compute at inference.
What makes K3 more than a bigger K2 is a pair of architectural changes:
- Kimi Delta Attention (KDA) — a hybrid linear attention mechanism. Moonshot says it delivers up to 6.3x faster decoding in million-token contexts, which is the difference between a 1M window being a spec-sheet number and an actually usable one.
- Attention Residuals (AttnRes) — a drop-in replacement for standard residual connections that retrieves representations selectively across depth. Moonshot reports roughly 25% higher training efficiency at under 2% additional cost.
Stack those on top of refined data and training recipes and the company claims about 2.5x better overall scaling efficiency than Kimi K2. Serving was clearly a design constraint, not an afterthought: K3 uses quantization-aware training from the SFT stage onward, ships MXFP4 weights with MXFP8 activations for broad hardware support, and — because KDA breaks naive prefix caching — Moonshot contributed its own implementation to vLLM.
The benchmarks: leading in the messy places
Moonshot is refreshingly direct about where K3 sits. Overall, it still trails the two strongest proprietary models, Claude Fable 5 and GPT-5.6 Sol. But on its own evaluation suite, K3 tops the field on several benchmarks that matter for real agentic work.
| Benchmark | Kimi K3 | Fable 5 | GPT-5.6 Sol | Opus 4.8 |
|---|---|---|---|---|
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 |
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 |
The pattern is telling. K3 leads on long-horizon, multi-step, and document-heavy tasks — SWE Marathon, BrowseComp, OmniDocBench — while trailing Fable 5 on FrontierSWE and GPT-5.6 Sol on DeepSWE. In other words, it's strongest exactly where a 1M-token window and native multimodality pull their weight.
One caveat worth knowing: the BrowseComp result used context compaction triggered at 300K tokens. Without that context management, K3 scores 90.4 — still frontier-class, just not the headline 91.2.
Multimodal by design, not by bolt-on
K3 handles text, images, and video through one native multimodal architecture rather than a vision encoder stapled onto a language model. Moonshot's reported use cases lean into this: iterating between code and live screenshots, parsing complex documents (that 91.1 on OmniDocBench), and reproducing research papers end to end. One example cites a session that pulled from 20-plus papers and produced 3,000-plus lines of Python inside a single context window.
What it costs
Pricing is flat, with no tiering by context length:
- Cache-hit input: $0.30 / million tokens
- Cache-miss input: $3.00 / million tokens
- Output: $15.00 / million tokens
The 10x gap between cache-hit and cache-miss input makes the cache-hit rate the number to watch. Moonshot reports above 90% cache hits in coding workloads, where the same repository context gets reused across many turns — which is where the pricing actually gets cheap.
Access runs through the OpenAI SDK against a Moonshot base URL, so migration is mostly a config change:
from openai import OpenAI
import os
client = OpenAI(api_key=os.environ["MOONSHOT_API_KEY"],
base_url="https://api.moonshot.ai/v1")
completion = client.chat.completions.create(
model="kimi-k3",
reasoning_effort="max",
messages=[{"role": "user", "content": "Introduce Kimi K3 in one sentence."}],
)
print(completion.choices[0].message.content)
A few rules trip people up: reasoning_effort only accepts max, the old K2.x thinking parameter is gone, and temperature, top_p, and n are fixed — omit them. max_completion_tokens defaults to 131,072 and can reach the full 1,048,576.
The Bottom Line
Kimi K3 is the strongest signal yet that open weights are trailing the frontier by weeks, not years. It doesn't beat Fable 5 or GPT-5.6 Sol outright, and Moonshot doesn't pretend otherwise. But a 2.8T open model that leads on long-horizon coding, agentic browsing, and document understanding — at $3 per million cache-miss input tokens and free to self-host after July 27 — reshapes the calculus for anyone who was paying frontier prices for those exact workloads. The interesting question now isn't whether open models can compete. It's what teams do with a frontier-class model they're allowed to download.
