Moonshot AI released Kimi K3 on July 16, 2026, and the headline number is hard to ignore: 2.8 trillion parameters, which the company says makes it the largest open-source model ever released. The weights are not public yet — they are scheduled for July 27 — but the model is already live in Kimi’s apps, in Kimi Code, and on OpenRouter. The scores are the real story.
On the Artificial Analysis Intelligence Index, K3 lands at 57. That puts it above Claude Opus 4.8 (56) and behind only GPT-5.6 Sol (59) and Claude Fable 5 (60) — the first time an open-weight model has been measured inside the frontier group rather than a tier below it. The twist is the price tag. K3 lists at $3.00 per million input tokens and $15.00 per million output, roughly three times what Kimi K2.6 charged. The era of frontier Chinese models at rock-bottom prices looks like it is ending. Here is what is real, what is vendor-reported, and where K3 actually fits.
Key takeaways
- 2.8T parameters, 16 of 896 experts active. A sparse Mixture-of-Experts built on Moonshot’s “Stable LatentMoE” framework — the largest open model announced to date.
- 57 on the AA Intelligence Index — above Claude Opus 4.8 (56), below GPT-5.6 Sol (59) and Claude Fable 5 (60). The strongest open-weight score yet recorded.
- 1M context, native vision, thinking always on. Maximum thinking effort is the default; low- and high-effort modes are promised in later updates.
- Two new architecture pieces: Kimi Delta Attention (up to 6.3× faster decoding at million-token contexts) and Attention Residuals (~25% better training efficiency for under 2% extra cost).
- Not cheap anymore. $3.00 in / $15.00 out per 1M tokens ($0.30 on cache hits) — about 3× K2.6’s $0.95/$4.00, and exactly Claude Sonnet 5’s list price.
- Weights due July 27, 2026. Until then it is API-only — and at roughly 1.4 TB in 4-bit, “open” will not mean “runnable” for almost anyone.
What Kimi K3 actually is
K3 is Moonshot’s general flagship, not a specialist. That is a deliberate change of direction from Kimi K2.7 Code, which the company split off as a coding-only release a month earlier. K3 is built to do everything: chat, long-document work, vision, and — the part Moonshot clearly cares most about — long-horizon agentic tasks where a model plans, calls tools, reads results, and keeps going.
The scale is the first thing to understand, and the sparsity is the second. Of the 2.8 trillion total parameters, only 16 of 896 experts fire on any given token. That is what keeps inference cost and latency in a range an API can actually serve; a dense 2.8T model would be economically impossible to run. The trade is memory: every one of those 2.8 trillion parameters still has to sit in VRAM, whether it activates or not.
| Spec | Kimi K3 |
|---|---|
| Developer | Moonshot AI |
| Total parameters | 2.8 trillion (MoE) |
| Active per token | 16 of 896 experts |
| Context window | 1M tokens |
| Modality | Text, vision → text |
| Reasoning | Always on (max effort by default) |
| Quantization | MXFP4 weights, MXFP8 activations |
| Input price | $3.00 / 1M ($0.30 on cache hit) |
| Output price | $15.00 / 1M |
| Output speed | ~62 tokens/sec |
| Released | July 16, 2026 |
| Open weights | Scheduled July 27, 2026 |
Full specs and live pricing sit on the Kimi K3 spec sheet in our AI models database.
The architecture: how you train 2.8T without the bill exploding
Two pieces of Moonshot’s own research carry this release, and both target the same problem — that scaling a transformer normally means paying more for every extra token of context and every extra layer of depth.
Kimi Delta Attention (KDA) is a hybrid linear attention mechanism. Standard attention costs grow quadratically with sequence length, which is exactly why million-token contexts have been slow and expensive everywhere they have shipped. Moonshot reports KDA delivers up to 6.3× faster decoding at million-token contexts — the difference between a 1M window that exists on a spec sheet and one you would actually use.
Attention Residuals (AttnRes) is described as a drop-in replacement for standard residual connections, improving how signal flows through depth. Moonshot reports roughly 25% higher training efficiency for under 2% additional cost. Together with the Stable LatentMoE framework, Gated MLA and a new activation (SiTU), the company claims about a 2.5× improvement in overall scaling efficiency versus Kimi K2.
Those efficiency numbers are vendor-reported and not yet independently reproduced. But they explain the strategy: you do not get to 2.8T by buying more GPUs than Google — export controls make that route unavailable to a Chinese lab. You get there by making each GPU-hour go further.
Benchmarks: where it wins, where it doesn’t
K3’s strongest results cluster around agentic and reasoning work rather than raw conversation.
| Benchmark | Kimi K3 | What it measures |
|---|---|---|
| GPQA Diamond | 93.5% | Graduate-level science reasoning — strongest open-weight result published at release |
| BrowseComp | 91.2% | Web research agents — best published score on the tracker at release |
| Terminal-Bench 2.1 | 88.3% | Command-line / shell agent tasks |
| MCP Atlas | 84.2% | Tool use over Model Context Protocol |
| MMMU-Pro | 81.6% | Multimodal understanding |
| DeepSWE | 67.5 | Software engineering |
| Humanity’s Last Exam (with tools) | 56.0% | Hardest general reasoning set |
| AA Intelligence Index | 57 | Composite — #4 of 189 models tracked |
Two independent signals stand out. In blind testing on Arena, developers preferred Kimi over every leading US model for front-end coding — including Fable 5 and GPT-5.6 Sol. And on real-world task automation, K3 ranked first in four of eight benchmarks (including Automation Bench, SpreadsheetBench 2 and BrowseComp), finishing second to Fable 5 in most of the rest.
The honest summary: K3 still trails Fable 5 and GPT-5.6 Sol overall, and it beats essentially everything else that has been measured. For an open model, that has never been true before.
The price story: the cheap-Chinese-AI era is ending
This is the part that gets less coverage and matters more. Chinese labs built their reputation on undercutting Western APIs by an order of magnitude. K3 does not do that.
| Model | Input / 1M | Output / 1M | Cache hit |
|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 |
| Kimi K2.6 (predecessor) | $0.95 | $4.00 | $0.16 |
| Claude Sonnet 5 | $3.00 | $15.00 | — |
| Claude Opus 4.8 | $5.00 | $25.00 | — |
| GPT-5.6 Sol | $0.50 | $30.00 | — |
K3 costs roughly three times its own predecessor and lists at exactly Claude Sonnet 5’s price. On a per-task basis the gap narrows further: measured averages put K3 at about $0.94 per task, GPT-5.6 Sol at $1.04 and Opus 4.8 at $1.80. K3 is still cheaper — but it is now competing on value at the margins, not on being 10× cheaper. Frontier-grade reasoning appears to cost roughly what it costs, regardless of who trains it.
Intelligence per dollar: our take
Raw price is the wrong metric. What matters is how much capability each dollar buys. Using the blended price and intelligence scores from our 2026 AI Price-Performance Index, here is where K3 lands:
| Model | Intelligence | Blended $/1M | Intelligence per $ |
|---|---|---|---|
| Kimi K3 | 57 | $9.00 | 6.3 |
| Claude Opus 4.8 | 55.7 | $15.00 | 3.7 |
| GLM 5.2 | 51.1 | $2.90 | 17.6 |
| DeepSeek V4-Flash | 40.3 | $0.21 | 192 |
Three conclusions fall out of that table. K3 delivers about 1.7× the intelligence per dollar of Claude Opus 4.8 while scoring slightly higher — a genuinely better deal at the top end. But GLM 5.2 still returns 2.8× more capability per dollar than K3 at six points lower intelligence, and DeepSeek V4-Flash returns about 30× more. K3 is the smartest open model available; it is nowhere near the best value one. If you are paying frontier prices, you should be certain you need frontier reasoning. Run your own numbers in the AI API cost calculator, or see the full ranking on the LLM leaderboard.
“Open weights” does not mean you can run it
When the weights land on July 27, expect a wave of headlines about anyone being able to run a frontier model at home. Check the arithmetic first.
At 2.8 trillion parameters in native 4-bit (MXFP4), the weights alone come to roughly 1.4 TB. Add a KV cache sized for anything near the 1M context and you need more. Realistically that is on the order of 16 H200-class GPUs across two nodes — several hundred thousand dollars of hardware before power and networking. For comparison, K2.7 Code at 1T needed about 595 GB and eight 80GB GPUs, and that was already out of reach for individuals.
So who is the weight release actually for? Sovereign deployments, regulated enterprises that cannot send data to an API, research labs, and cloud providers who will host it for everyone else. That is still a meaningful gap versus a closed model — you can audit it, fine-tune it, and run it inside your own walls — but it is not a home GPU story. If you want to know what your hardware can actually hold, our LLM VRAM calculator does the math per model, and the self-hosting vs API calculator shows where owning GPUs starts to beat paying per token.
Who should use K3
Use it if you are running agentic workloads — browser automation, multi-step tool chains, long-horizon coding — where its BrowseComp, Terminal-Bench and MCP Atlas scores translate into fewer failed runs. It is also the obvious pick if you want frontier-class reasoning with a credible path to self-hosting later, or if front-end code quality matters (developers picked it over Fable 5 blind).
Skip it if your work is ordinary chat, summarization, classification or retrieval. At $3/$15 you would be paying frontier rates for tasks that GLM 5.2 or DeepSeek V4-Flash handle at a fraction of the cost. And skip it if you assumed “open” meant you could download it this week — the weights are still nine days out at the time of writing, and 1.4 TB when they arrive.
The larger point is the one the benchmark table makes quietly. An open-weight model just posted a score above Claude Opus 4.8. Whatever gap existed between open and closed frontier AI is now measured in a couple of index points and a few months — not in generations.
FAQ
Is Kimi K3 better than Claude Opus 4.8?
On the Artificial Analysis Intelligence Index, yes — K3 scores 57 against Opus 4.8’s 56, and it costs $3/$15 per million tokens versus Opus at $5/$25. It still trails GPT-5.6 Sol (59) and Claude Fable 5 (60).
Is Kimi K3 open source?
The weights are scheduled for public release on July 27, 2026, following Moonshot’s Modified MIT precedent with earlier Kimi models. Until then K3 is API-only through Kimi’s apps, Kimi Code and OpenRouter.
How much does Kimi K3 cost?
$3.00 per million input tokens, $15.00 per million output, and $0.30 per million on cache hits. That is about 3× the price of Kimi K2.6 and identical to Claude Sonnet 5’s list price.
Can I run Kimi K3 locally?
Almost certainly not. At 2.8 trillion parameters the 4-bit weights are roughly 1.4 TB — around 16 H200-class GPUs across two nodes, before any KV cache for its 1M context. It is a data-centre model, not a desktop one.
How big is Kimi K3?
2.8 trillion total parameters in a Mixture-of-Experts design, with only 16 of 896 experts active per token. Moonshot says that makes it the largest open-source model released to date.
