Kimi K3 · Moonshot AI · Infrastructure · 2026
The file weighs 1.56 TB. Downloading it is free. Turning it on needs the most expensive hardware on the planet. AWS, Google Cloud and Azure received the same model the same weekend — and each response was a portrait of their infrastructure.
The paradox
On the last Sunday of July, Moonshot, a Beijing-based lab, released Kimi K3's weights: 2.8 trillion parameters, history's largest open model, free to download on Hugging Face. And in the following three days, AWS, Google Cloud and Microsoft Azure moved to publish support. Why the rush over something free? Because downloading it is free, but running it isn't — and in that difference lies one of the biggest businesses of the decade.
The protagonist, with just what you'll need afterward.
Mixture of experts: on each token, only a fraction of the model works. It's like having 896 specialists in the room, but consulting 16 per question.
The compute per token is much lower than a dense 2.8-trillion model — but the memory is still that of the full 2.8 trillion, because every expert has to be loaded and ready across the GPUs.
Moonshot AI → Beijing-based lab
Architecture → MoE (mixture of experts)
Parameters → 2.8 trillion total, 16 of 896 experts active per token
Class → frontier — matches the best closed models on several
benchmarks, though not all of them"Releasing the weights" is the news. But precision matters here — words are used as they should be.
With OpenAI or Anthropic: the binary does NOT exist for you.
The model lives at their place, you access via API.
With Moonshot: it published the full binary (the weights) on
Hugging Face. With the binary in hand,
you run it wherever you want.The weights download for free, under the Kimi K3 License (a named, specific license, not a generic one). Commercial use is allowed for a normal business.
The training recipe or the data it was trained on. And if you build a business serving the model above a certain size, or your product surpasses 100 million users, special clauses apply.
With all that, the precise answer is: it's not open source outright, it's open-weight. An open model at this level still changes the board — you decide where it runs, what it sees, and your prompts never leave your perimeter. It's not total sovereignty, because you still depend on the cloud and everything around the model, but control changes hands.
The downloaded checkpoint is ~1.56 TB. Serving it — runtime, buffers, context memory — requires over 1.5 TB of GPU VRAM, not disk space. From least to most:
Your laptop
Not even close. Forget it.
512 GB Mac Studio (the priciest one you can buy)
Loads a third of the checkpoint.
An 8× H100 node (last year's gold standard)
Still not close.
An 8× B200 node (the new Blackwell)
Still not enough.
An 8× B300 node (Blackwell Ultra)
Barely fits.
Not my own math
This is the official vLLM guide, the inference engine that serves this model: "the entire model can barely fit in a single NVIDIA DGX B300 and requires a minimum of 16 NVIDIA B200/GB200 GPUs to serve on that hardware generation." A single node only fits 8 B300s, period. Kimi K3 is, in practice, a free model that needs the most powerful hardware on the planet to turn on — and that's the answer to why the three cloud giants rushed to publish guides the same week: every company that wants this in-theory-free model is going to need to rent the type of infrastructure they sell. The open model doesn't compete with them, it sells them hardware.
When the three giants receive the same model the same weekend, each response is a portrait of the hardware they have and the strategy they play.
Hardware
8× B300, one node
Engine
vLLM (TP8)
Orchestration
SageMaker HyperPod / EKS
Price
$112.32/h (Capacity Blocks, official)
You control everything, we sell you the densest node possible. The price is on the shelf — but it's reserved, not improvised.
Hardware
16× GB200, 4 coordinated nodes
Engine
SGLang (TP16, multi-host)
Orchestration
GKE / Model Garden
Price
Not public (preview, no committed-use discounts)
More total memory, more headroom for a million-token context. The cost is complexity: coordinating 4 nodes is much harder than one.
Hardware
Abstracted — you never see the GPUs
Engine
Not exposed
Orchestration
Microsoft Foundry
Price
Public, per token (via Fireworks AI)
You operate nothing and pay per token, in exchange you give up control of the runtime. Azure does have Blackwell available — this was a product decision, not a lack of hardware.
AWS GOOGLE AZURE (Foundry)
Strategy Dense Distributed or Managed inference
self-hosting 1-click (Model Garden) (Fireworks)
Hardware 8× B300, 1 node 16× GB200, 4 nodes Abstracted
Engine vLLM (TP8) SGLang (TP16) Not exposed
Orchestration HyperPod / EKS GKE / Model Garden Foundry
Purchase Reserved Quota + preview Per token
capacity (sales)
Infra control High High or medium Low
Weight control Yours Yours No
Price $112.32/h Ask sales Public, per token
(Capacity Block) (high estimate)AWS: $112.32 an hour is around $2,700 a day — leave it on guard the full month and the bill would pass $80,000/month, for a single model. Google doesn't publish the price for its recommended configuration: it's in preview, no committed-use discounts, and the number is discussed with sales — based on GB200 market prices elsewhere, it's estimated well above AWS, though that's just an estimate. Azure chose not to expose the hardware: it has had Blackwell available since late 2025 (NVLink domains up to 72 GPUs), so it's not a lack of hardware — it was a product decision to move fast with an inference specialist inside its catalog.
The point of an open model is that you don't depend on a single provider. Someone already serves it and lets you try it for free.
Try it free
A serverless compute platform. It already serves Kimi K3 with an OpenAI-compatible API, and its plan includes $30 of free compute per month — plenty to try the model.
Use it in production
$3 per million input tokens ($0.30 on a cache hit) and $15 per million output tokens. No infrastructure of your own.
The second half of the paradox
What you can try for free in your terminal is the same 2.8-trillion-parameter model that needs 8 B300s or 16 B200s to self-host — because obviously someone else put up the data center. The open model created an entire market of inference middlemen: Fireworks serves it to Microsoft, Modal serves it to you, and even Moonshot sells it to you by API.
In the previous video I built a war room where two rival companies' agents, Claude (Anthropic) and Codex (OpenAI), operate my infrastructure with cryptographic identity.
Already there
Claude · Anthropic
Codex · OpenAI
The third joins
Kimi K3 · Moonshot
Same conditions
Its own key. Its own cage. The same human approval.
Since the room runs on an open protocol, any agent backed by any model can join — with the rules I defined. Three companies, OpenAI, Anthropic and Moonshot, with three models and three different cryptographic passports, sitting in the same room, mine. The cloud giants rushed this week to host the Chinese model. I have it on standby right here, and it cost me nothing. Watch how the room was built in Buzz ›
A Chinese lab released a frontier-class model for free, and the three cloud giants reordered their week to host it. According to reports, Microsoft is even said to be evaluating it for Copilot.
The thesis
When weights become portable and APIs start to look alike, power moves from the lab toward the infrastructure that serves the model: the memory, the network, the inference engine, the available capacity, and the governance around it. Weights can be free, but operating them never is. That's why the clouds run — the free model sells them hardware — and that's why on this channel we build systems, not demos: the value is no longer in accessing the model, which keeps getting more open and cheaper, but in the architecture around it. The identities, the cages, the credentials, the orchestration — nobody gives you that for free.
The video's official sources, verified.
AWS (official blog) · July 30, 2026
AWS's official recipe: a P6-B300 instance with 8 B300 GPUs, vLLM as the engine, orchestrated with HyperPod or EKS.
Google Cloud (official forum) · July 27, 2026
The fastest support of the three clouds: one-click Model Garden, or the official AI Hypercomputer recipe for GKE.
Microsoft (Azure AI Foundry Blog) · July 28, 2026
Azure's announcement: Kimi K3 served by Fireworks AI inside the Foundry enterprise catalog, no GPU provisioning required.
vLLM Project · July 27, 2026
"The entire model can barely fit in a single NVIDIA DGX B300 and requires a minimum of 16 NVIDIA B200/GB200 GPUs to serve on that hardware generation." The technical source for the VRAM ladder.
Moonshot AI · 2026
The official Hugging Face repo: the full weights (1.56 TB, 96 shards) and the Kimi K3 License that comes with them.
Moonshot AI (platform.kimi.ai) · 2026
$3.00 per million input tokens (cache miss), $0.30 on a cache hit, $15.00 per million output tokens.
The agent room, multi-agent infrastructure, and the discipline behind all of it.
Where Kimi enters as a third agent: the war room with cryptographic identity where Claude and Codex were already running my infrastructure.
Another real multi-agent infrastructure case: 7 independent services, each with its own URL and its own container.
The discipline behind all of this: the model barely matters anymore — what matters is the environment and controls you build around it.
The essentials on Kimi K3.
It's Moonshot AI's flagship model, a lab based in Beijing. It uses a MoE (mixture of experts) architecture with 2.8 trillion total parameters, of which only a fraction works on each token — like having 896 specialists in the room but consulting 16 per question. On several benchmarks it reaches the level of the best closed models, though it still trails on others. It's, today, the most capable open model you can download.
Downloading it is free: Moonshot published the full weights on Hugging Face, at no cost. But running it isn't free — it needs over 1.5 TB of VRAM, hardware only a handful of nodes on the planet have. That's the paradox at the center of the video: the model is free, the infrastructure to turn it on is the most expensive thing that exists. If you don't want to pay for that infrastructure, you can try it free with Modal (no GPU needed) or pay per token via API.
It's open-weight, not open source outright. Moonshot publishes the trained binary (the weights) under the Kimi K3 License: they download for free and commercial use is allowed, but with conditions for large players — if you build a business serving the model above a certain size, or your product surpasses 100 million users, special clauses apply. What you don't get is the training recipe or the data it was trained on, which is what "open source" strictly means.
The downloaded checkpoint is ~1.56 TB, and serving it (runtime, buffers, context memory) requires over 1.5 TB of GPU VRAM, not disk space. According to the official vLLM guide, the full model barely fits on a single 8-GPU B300 (Blackwell Ultra) node and requires a minimum of 16 B200/GB200 GPUs to serve on the previous hardware generation. An 8× H100 node — last year's standard — isn't even close to enough.
AWS publishes its price: $112.32 per hour on Capacity Blocks for the 8-GPU B300 instance — around $2,700 a day, over $80,000 a month if you leave it running 24/7. Google Cloud doesn't publish the price for its 16-GPU GB200 configuration (it's in preview, you talk to sales); based on market prices, it's estimated well above AWS, though that's just an estimate. Azure doesn't charge you per GPU-hour: you pay per token via Fireworks AI, without touching any hardware.
Modal, a serverless compute platform, already serves Kimi K3 with an OpenAI-compatible API, and its plan includes $30 of free compute per month — plenty to try the model. If you want to use it in production without your own infrastructure, Moonshot's official API charges $3 per million input tokens (or $0.30 on a cache hit) and $15 per million output tokens.
It's the Beijing-based Chinese lab behind Kimi K3. On July 26, 2026 it released the model's weights, and in the following three days AWS, Google Cloud and Microsoft Azure published official support — Google with day-0, Microsoft adding it to its enterprise catalog, and AWS with its deployment guide. According to press reports (unconfirmed officially), Microsoft is said to be evaluating Kimi K3 for some Copilot workloads.
Community
The full analysis, the comparison tables for the three clouds, and how the agent room with Kimi, Claude and Codex was built — step by step. Free access to the community; full courses live in the Premium tier.
Join Agentic Engineers →YouTube channel
@NicolasNeiraGarcia
ADK · A2A · Claude Code · Automation · Infrastructure