Kimi K3 is here: a 2.8-trillion-parameter open model just hit the frontier
Moonshot AI dropped Kimi K3 overnight — 2.8T parameters, a 1M-token context, open weights — and it is already topping benchmark boards. Here is every number, with the caveats.
Moonshot AI shipped Kimi K3 on 16 July 2026 with almost no ceremony — a model card, an API endpoint, and open weights. Within hours it was sitting at #1 on several public benchmark boards. I pulled every number I could find from the model card and the trackers so you don't have to. Here's the whole picture, with the caveats where they belong.
What Moonshot actually shipped
Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model — by Moonshot's own claim, the largest open-weight model released to date. It's natively multimodal (text, images, video in), with a full 1,048,576-token context window at a flat price — no tiering by context length.
The architecture is the interesting part. K3 runs on Kimi Delta Attention (KDA) — a hybrid linear-attention mechanism — plus Attention Residuals, which is how they keep a 1M-token window economical instead of quadratically expensive. That's not a marketing detail; it's the reason the pricing below is even possible.
It ships open-weight, so it's not just an API — you can (in principle) self-host it. In practice, "self-host a 2.8T model" is a data-center sentence, not a laptop one. For almost everyone, the API is the product.
The benchmarks
Moonshot published K3's scores across four tracks — agentic, coding, knowledge, and multimodal. Start with the headline numbers, because these are the ones doing the talking:
A 93.5 on GPQA-Diamond — PhD-level science questions — is frontier-tier, full stop. And a 88.3 on Terminal-Bench 2.0 (can the model actually operate a shell to finish real tasks) is the number that should make anyone building coding agents look up.
Coding
| Benchmark | Kimi K3 |
|---|---|
| Terminal-Bench 2.0 | 88.3 |
| FrontierSWE | 81.2 |
| ProgramBench | 77.8 |
| Kimi Code Bench v2 | 72.9 |
| DeepSWE | 67.5 |
| MLS-Bench Lite | 48.3 |
| SWE Marathon | 42.0 |
| PostTrain Bench | 36.6 |
Agentic
| Benchmark | Kimi K3 |
|---|---|
| DeepSearchQA (F1) | 95.0 |
| BrowseComp | 91.2 |
| MCP Atlas | 84.2 |
| DECK-Bench | 73.5 |
| Toolathlon-Verified | 73.2 |
| JobBench | 52.9 |
| APEX-Agents | 37.6 |
| SpreadsheetBench 2 | 34.8 |
| AutomationBench | 30.8 |
| GDPval-AA / AA Briefcase | 1668 / 1548 Elo |
Knowledge & reasoning
| Benchmark | Kimi K3 |
|---|---|
| GPQA-Diamond | 93.5 |
| Humanity's Last Exam (with tools) | 56.0 |
| Humanity's Last Exam (no tools) | 43.5 |
| Chatbot Arena (text, overall) | 1486 Elo |
Multimodal
| Benchmark | Kimi K3 |
|---|---|
| MathVision (with Python) | 97.8 |
| MathVision | 94.3 |
| CharXiv Reasoning | 91.3 |
| OmniDocBench | 91.1 |
| BabyVision (with Python) | 85.7 |
| CharXiv (no tools) | 84.8 |
| MMMU-Pro (with Python) | 83.4 |
| MMMU-Pro | 81.6 |
| OfficeQA Pro | 63.3 |
| PerceptionBench | 58.5 |
| WorldVQA (force-answer) | 51.0 |
| ZeroBench (pass@5, with Python) | 41.0 |
How it stacks up
Independent, apples-to-apples leaderboards are still catching up — most of the numbers above are Moonshot-reported. But the early read from people who've run it is consistent: K3 matches or beats Opus 4.8 and GPT-5.5 on several coding benchmarks, and its overall quality lands between GPT-5.6 and Claude Fable 5, with visual reasoning that edges past Fable 5. TechCrunch framed the release as Moonshot finally "closing the gap" with the closed frontier. On these numbers, "closing" is underselling it.
The part that actually reorders the market is price:
| Model | Input $/1M | Output $/1M | Context | Open weights |
|---|---|---|---|---|
| Kimi K3 | $3 (cached $0.30) | $15 | 1M | Yes |
| Claude Opus 4.8 | $5 | $25 | 1M | No |
| Claude Fable 5 | $10 | $50 | 1M | No |
Frontier-adjacent quality, a 1M-token window, and open weights, at roughly a third of a closed flagship's output price. That combination is why this launch matters more than the leaderboard screenshots.
The honest caveats
I like this model, and I'm still going to say the quiet part:
- Most scores are first-party. Moonshot published them; independent replications are landing now, not settled. Treat the table as "vendor-reported, promising" until the neutral boards agree.
- Open weight ≠ runnable. 2.8T parameters is open in the licensing sense; hosting it is a serious infrastructure commitment. For nearly everyone this is an API model.
- Benchmarks aren't your workload. GPQA 93.5 is thrilling; whether K3 is your best model depends on your prompts, your latency budget, and your tolerance for a Chinese-lab dependency in your stack.
My take
I build for portability on purpose — I don't want any one lab holding my roadmap. Kimi K3 is exactly why that stance pays off: a genuinely frontier-class option, open-weight, priced to undercut, dropped overnight. Wire it behind an interface, run it against your own evals, and let it compete for your traffic on merit. If the independent numbers hold even most of the way, the "closed models are just better" argument got a lot weaker this week.
Sources
- Moonshot AI — Kimi K3 quickstart / platform docs and OpenRouter model card
- Benchmark suite — BenchLM: Kimi K3
- TechCrunch — Moonshot's Kimi 3 expected to close the gap with Opus 4.8
- BigGo — Moonshot launches flagship K3, coding benchmarks rival GPT-5.5
- OfficeChai — K3: largest open model, 2.8T params, 1M context, pricing
Numbers are as reported on launch day, 16 July 2026, and will move as independent benchmarks land. If you spot a corrected figure, tell me and I'll update.
