On July 27 2026 Moonshot dropped Kimi K3 weights on Hugging Face. 2.8 trillion parameters. Open license. First open model that trades blows with GPT-5.6 and Claude Fable 5 on every major benchmark.

This is not another incremental release. This is the point where the entire industry flipped. For the first time, frontier intelligence is not something you rent by the token. It is something you download, run on your own hardware, modify, and redistribute.

The K3 numbers that matter

Everyone is talking about the 2.8T total parameter count. That is the wrong number.

K3 is a sparse MoE. Only 104B parameters activate per token. That is the number that matters for inference cost. This model was not built to be big. It was built to be cheap to run.

Metric	Value
Total parameters	2.8T
Activated per token	104B
Experts total	896
Experts active	16
Context window	1,048,576 tokens
Quantized weight size	1.4 TB
Minimum GPU nodes (A100 80GB)	3
Minimum GPU nodes (H200)	2
Minimum GPU nodes (B300)	1

It was trained with quantization aware training for MXFP4. This is not an after the fact quant. This model was built to run compressed. That is the single most important architectural choice made here.

Deployment math: what it actually costs to run frontier

Everyone posting memes about running K3 on a gaming PC is wrong. But everyone saying you need a datacenter is also wrong.

We have real preliminary numbers from operators that stood the model up within 12 hours of release:

On B300 hardware you can run K3 for $3.18 per million output tokens. That is less than one fifth the cost of the equivalent closed API. Even on obsolete A100s you undercut OpenAI's pricing.

This is the number that broke the market. This is why Zhipu lost 47% of its market cap in two trading days. This is why semiconductor stocks moved. No one had run the numbers correctly before release.

The benchmark dead heat

K3 does not win every benchmark. It does not need to. It is within 3% of the closed frontier models on 92% of published benchmarks. On coding and agent benchmarks it beats them.

Most importantly: it does this while being open. You can inspect every weight. You can fine tune it. You can run it air gapped. You can distill it. None of the closed models let you do any of this.

The Stack v3: the quiet foundation no one is talking about

Three days before K3 dropped, Hugging Face released The Stack v3. No press release. No blog post. Just 15.9 TB of cleaned, deduplicated, repository grouped source code. 4.9 trillion tokens. August 2025 cutoff.

This is the training dataset that every open code model will use for the next 18 months. Before this release, every team training a code model was crawling GitHub privately. Now everyone has exactly the same base corpus.

This is not an accident. This is infrastructure. This is how you commoditize the input side of model training.

Distillation is no longer theft. It is the industry standard.

On July 26 Jensen Huang posted on X that distillation is fundamental to intelligence. This was not an offhand comment. This was Nvidia formally declaring the end of the argument.

For two years closed model vendors have argued that using their outputs to train other models is theft. That argument is over. The entire industry has now lined up against it. Everyone except Anthropic.

Distillation datasets are now the fastest growing category on Hugging Face. In the last 30 days 117 separate distillation datasets were uploaded. 72 of them target K3.

Quantization did not break the model. It broke the assumptions.

We have operated for three years on the assumption that quantization is a dirty tradeoff. You trade quality for size. That assumption died last week.

Independent testing on Qwen 3.6 27B shows that 4 bit quantization produces results indistinguishable from full precision BF16 on every benchmark run. Even 3 bit quantizations retain >95% of performance if done correctly.

Quantization	AIME 120 Score	KL Divergence	Size relative to BF16
BF16	70.8%	0.00	100%
Q8_0	70.1%	0.012	50%
Q6_K	69.7%	0.027	37.5%
Q4_K_M	69.2%	0.041	25%
Q3_K_M	73.3%	0.048	18.75%
Q3_K_S	54.2%	0.092	18.75%

There is a cliff. But it is not at 8 bit. It is not at 4 bit. It is somewhere around 2.7 bit. No one saw this coming. All of our old rules were wrong.

The great American open model abdication

There is no US based frontier open model. Not one.

Meta abandoned open source when they pivoted Llama 4 to closed. OpenAI will never release weights. Anthropic is lobbying to ban open models above 70B parameters. Google releases obsolete models 12 months after they are no longer competitive.

This is not a failure of engineering. This is a failure of economics. US model companies have business models built on renting tokens. Open weights destroy that business model. They cannot compete on this playing field, so they will not play.

This is not the path to AGI. That is the point.

All of this progress will not get us AGI.

Ilya Sutskever is right. Scaling and distillation and benchmark optimization will not produce general intelligence. They will produce extremely capable, extremely cheap, extremely reliable tools. That is enough.

We are not building a ladder to the sky. We are paving the road right in front of us. And that road is already good enough to run an industrial revolution on.

The coming split

We are now entering a permanent two track industry.

One track will be closed models. They will run the most dangerous research. They will chase AGI. They will be expensive, regulated, controlled, and run by a tiny number of companies.

The other track will be open models. They will run everything else. Every API, every agent, every internal tool, every application. They will get cheaper every month. They will get faster every month. No one will control them.

No one is going to win this fight. Both tracks will exist. Both tracks will progress. They will feed each other. They will compete. That is the future.

What happens next

Over the next 90 days:

• K3 will be distilled down to 120B, 70B, 35B and 14B variants. The 35B variant will run on a single RTX 5090 and match GPT-5.5 performance.
• At least three other teams will release 2T+ open models before the end of the year.
• Every major cloud provider will launch hosted K3 instances at half the price of GPT-5.
• Anthropic will double down on lobbying for open model regulation.
• Someone will find the first critical security vulnerability in a frontier open model.

The frontier is open now. You cannot put that back in the bottle.

This is not the end of the race. This is the end of the first lap.