Kimi K3: The World's Largest Open-Weight AI Model – and Why "Open Source" Doesn't Quite Fit Here
2.8 trillion parameters, rank 4 worldwide – but "open source" is marketing here. What Kimi K3 can really do and where the catches are.
Kimi K3: The World's Largest Open-Weight AI Model – and Why "Open Source" Doesn't Quite Fit Here
In mid-July 2026, the Chinese AI lab Moonshot AI dropped a real bomb into the AI landscape: Kimi K3 – a model with 2.8 trillion parameters, native multimodal support, a context window of one million tokens, and benchmark results that leave prominent US models behind in several categories. And the whole thing is supposed to be "open source." Whether the label delivers on its promise – let's take a look.
2.8 Trillion Parameters – but What Does That Actually Mean?
If you swallow briefly at "2.8 trillion parameters": completely understandable. Kimi K3 is the largest open-weight model ever published – and the first model ever in the so-called "Open-3T class." For context: OpenAI never officially communicated GPT-4's exact parameter count; estimates range somewhere between 200 billion and just over a trillion. K3 clearly surpasses that in every scenario.
The trick is the Mixture-of-Experts architecture (MoE). K3 has 2.8 trillion parameters, but only activates 16 of 896 experts per processed token – that's just ~1.8 % of the total model capacity. So the model is nowhere near as compute-intensive as the raw parameter count suggests. Still, running it locally requires serious GPU infrastructure – we're not talking about a MacBook experiment here.
What makes Moonshot technically interesting are the new architectural building blocks introduced for the first time with K3. The most important ones:
- Kimi Delta Attention (KDA): Improves information flow across very long sequences – crucial with a one-million-token context window.
- Attention Residuals (AttnRes): Increases precision when retrieving representations during training – another building block for more reliable reasoning.
- Quantile Balancing: More efficient expert routing in the MoE setup, ensuring the load is better distributed across the 896 experts.
Moonshot states that this achieves a 2.5x scaling efficiency over the predecessor Kimi K2 – K3 gets significantly more out of the same training effort. On top of that comes native support for text, images, and video, and the 1-million-token context window comes without any compression tricks.
Open Source or Open Weight? The Crux
Moonshot AI markets Kimi K3 as "open source." The AI community – in particular representatives of the Open Source Initiative (OSI) – sees things far more soberly. And the distinction matters for anyone making deployment decisions.
What Moonshot actually releases: the model weights, starting July 27, 2026, under a modified MIT license. That is generous and useful.
What Moonshot does not release: the training code, the training data, and the training infrastructure. In other words, exactly the part that would make a model a genuine open-source project by OSI standards. You can use, adapt, and deploy K3 – but you cannot reproduce it.
"Open Weight, not Open Source – whoever gets the weights doesn't get the recipe."
This follows exactly the pattern of Meta's LLaMA series: a lot of openness in usage, little transparency in the creation process. Anyone expecting true reproducibility or full data transparency will be disappointed. Anyone wanting to freely deploy a powerful model gets exactly that.
Benchmarks: Where K3 Shines – and Where It Doesn't
Let's get to the discipline that always gets the loudest in the AI world: benchmarks. Kimi K3 delivers a mixed but overall impressive picture here.
The Highlights
K3 is particularly strong in coding and agentic reasoning. In the Frontend Code Arena benchmark, K3 takes 1st place with 1,679 points – beating Claude Fable 5, one of Anthropic's strongest proprietary models. According to Arena.ai, K3 wins in 6 of 7 frontend domains. On the BrowseComp benchmark, which tests complex, long-horizon web research, K3 reaches a state-of-the-art score of 91.2/100.
The coding-specific results in detail (nxcode.io):
| Benchmark | K3 Score | What is measured |
|---|---|---|
| Terminal-Bench 2.1 | 88.3 | Terminal tasks in an isolated environment |
| FrontierSWE | 81.2 | Extremely hard software engineering tasks |
| ProgramBench | 77.8 | Reconstructing programs from binary behavior |
| DeepSWE | 67.5 | Fixing real repository issues |
| SWE Marathon | 42.0 | Multi-hour full-project work |
The Big Picture
In the aggregated ranking from BenchLM.ai – with over 54 published benchmarks – K3 lands at 4th place out of 200 models with an overall score of 80.96/100. The current frontier models ahead of it: Anthropic's Claude Mythos 5 and Fable 5 as well as OpenAI's GPT-5.6 Sol – all proprietary, closed systems:
| Model | BenchLM Score |
|---|---|
| Claude Mythos 5 | 83.93 |
| Claude Fable 5 | 83.68 |
| GPT-5.6 Sol | 81.96 |
| Kimi K3 | 80.96 |
| Claude Opus 4.8 | 78.34 |
| Grok 4.5 | 76.72 |
4th place worldwide as an open-weight model. That is a real statement.
The Weaknesses
No fair look without the downsides. At 62 output tokens per second, K3 is noticeably slower than the model median of 71.1 tokens/s. Even more striking: during evaluations K3 produced around 130 million output tokens – more than twice the median of 63 million. The model is, to put it diplomatically, very chatty. Anyone using K3 intensively in production will notice it on the bill.
The Elephant in the Room: Distillation, Hype, and Benchmark Games
As shiny as the numbers are – there is a serious debate about how they came about.
Anthropic and other US labs accuse Chinese labs – including Moonshot, DeepSeek, and MiniMax – of practicing "industrial-scale model distillation." That means: capabilities from proprietary frontier models are distilled into their own models through mass querying and imitation – without paying for a license or doing the training work. That is a serious accusation, legally and ethically unresolved.
On top of that come discrepancies in the benchmark results: the numbers from Arena.ai and Moonshot's own figures partly diverge. Since the model weights won't be released until July 27, 2026, there is currently no way to independently reproduce them – you either believe Moonshot, or you wait.
Patrick Moorhead, founder of Moor Insights & Strategy and a well-known US tech analyst, advises sobriety and warns against another wave of "AI panic": the focus is too often on hype rather than sober assessment. Bank of America analysts, on the other hand, see K3 as evidence that pre-training scaling combined with architectural innovation brings substantial progress – even under hardware constraints. Both readings are not mutually exclusive.
Chinese Models Simply Make Less Drama
One point that often gets lost in the technical discussion – and is very concretely relevant for companies in Switzerland and Europe:
In recent months, Anthropic models like Fable 5 and Mythos have caused considerable turmoil. Due to US export controls, these models were temporarily completely blocked for users outside the US – including Switzerland. What that means in practice: companies that had built on these models suddenly found themselves without access. We experienced this firsthand at Evoya and reported on the block and the later lifting of the export controls.
Kimi K3, on the other hand? Just there. No export restrictions, no drama, no weeks of waiting. This is not a political statement – it is a practical observation for companies that rely on AI models.
Mozilla CTO Raffi Krikorian put it succinctly: US labs are "evidently concerned" about the performance of Chinese models – which leads to an interesting dynamic: the more seriously the threat is perceived, the harder it becomes politically to simply ban open-weight models from China. The situation in the US remains dynamic: lawmakers are considering restrictions for US companies that use Chinese AI models. For European and Swiss firms this does not apply so far – but it pays to keep an eye on it.
As for price: at $15 per million output tokens, K3 sits where Claude Sonnet is – Chinese top models have thus arrived in the price region of Western mid-range models. That was different a year ago. Competition works.
What Does Kimi K3 Mean for Evoya Customers?
We at Evoya AI are currently testing Kimi K3 intensively and will integrate the model into the Evoya AI Workspace in the coming days. Especially in coding-heavy workflows and complex agentic tasks – exactly where K3 shines in the benchmarks – we want to see how it performs in real use.
But the real point is a different one: Evoya is built for exactly such moments.
When a model gets blocked due to export controls, when a new model like Kimi K3 shakes up the leaderboards, or when a provider suddenly raises prices – our customers don't have to worry about it. In the Evoya AI Workspace, all major models from all relevant providers are united on one platform: OpenAI, Anthropic, Google, Meta, Mistral – and soon Moonshot too. You switch the model with one click, compare results side by side, and are never dependent on a single provider.
No lock-in. No drama. Just the best model for the use case at hand.
Conclusion
Kimi K3 is real, it is impressive, and it deserves to be taken seriously. 4th place worldwide as an open-weight model, superior coding performance in several categories, a 1-million-token context window without compromises – this is not a hype model, it is a serious contender. The benchmark numbers should be read with healthy skepticism until the weights are out on July 27 and independent evaluations follow.
What can be stated clearly: the label "open source" is marketing here. "Open weight" is the correct term – useful, but not the same thing. And the question of which model fits best for which use case remains complex – which in turn argues for a platform that unites all relevant models.
Evoya AI is a Swiss AI agency based in Uster. In the Evoya AI Workspace, we unite all major AI models on one platform – privacy-compliant, flexible, and without provider lock-in.
