Today is Moonshot AI's own deadline for releasing the weights to Kimi K3, and as of this writing they are not out. The company's blog has said since launch that full weights would be published by July 27, 2026. Checked this morning, huggingface.co/moonshotai lists eighteen repositories and the newest is Kimi-K2.7-Code. There is no K3 repo. The day is not over in every timezone, so this may well resolve within hours, and the more interesting question is what happens if it does.

Because the number that matters is not the release date. It is 1.4 terabytes.

RelatedLaguna S 2.1: 8B Active Params, 70% Terminal-Bench

Where does 1.4 terabytes come from?

K3 is a 2.8-trillion-parameter mixture-of-experts model that activates roughly 50 billion parameters per token, routing through 16 of 896 experts. Mixture-of-experts is what makes it cheap to run per token. It does nothing at all to make it cheap to hold. Every one of those 2.8 trillion parameters has to be resident in memory, because any token might route to any expert.

At four-bit MXFP4 precision that is about 1.4 TB. At bf16, the precision most labs train and release at, it is around 5.6 TB. Those are floors, not estimates. They cover weights only, before the key-value cache for a 1-million-token context window, before activation memory, before running more than one request at a time.

Kimi K3 weight size against available accelerator memory Bar chart comparing Kimi K3 at bf16 requiring 5600 gigabytes and at MXFP4 requiring 1400 gigabytes against a single 80 gigabyte accelerator and an eight-card node totalling 1536 gigabytes. MEMORY FOOTPRINT · GB Weights only, before context or concurrency 5,600 K3 at bf16 1,400 K3 at MXFP4 1,536 8 x 192GB node 80 one 80GB card 2.8T params x precision. Arithmetic, not benchmark. genztech.blog
Fig 1 · footprint A full eight-card node of the newest silicon clears the four-bit version by 136 GB. That headroom has to cover a 1M-token context window.

In practical terms you need roughly eighteen 80GB accelerators just to load the four-bit version, or a single node of the newest silicon at eight cards of 192GB each, which fits the weights with almost nothing to spare. Blackwell-class NVIDIA parts and AMD's MI400 line are the realistic targets. This is not a home lab release, and it is not a two-3090s-and-a-dream release either.

Why publish weights nobody can host?

Three reasons, and only one of them is about you running the model.

The first is auditability. Anyone with the weights can inspect what the model actually is, run their own evaluations without going through Moonshot's API, and check claims independently. That matters more than usual for K3, given that the White House accused Moonshot on July 22 of distilling Anthropic's Fable to build it. Published weights are the only artifact that makes that argument checkable by third parties rather than a matter of assertion.

The second is that clouds can host it. Together, Fireworks, Baseten and similar providers do have the hardware, and once weights are public they compete on serving the same model. That is what has historically dragged the price of open-weight models below the original lab's own API within weeks.

The third is derivative work. Distillation, quantization to formats that do fit smaller hardware, fine-tunes on narrow domains. Nobody runs 2.8T at home, but a 30B distilled from it is a different conversation.

What does it cost to just rent it instead?

Moonshot's own pricing is $3 per million input tokens and $15 per million output, with cache hits at $0.30. On paper that is far cheaper than the $5 and $25 Anthropic charges for Claude Opus 5, and it is why K3 keeps getting described as the cheap route to a frontier score.

Our own reporting complicates that. When Fireworks AI ran K3 and Claude Fable 5 through roughly 1,030 tasks in real agent loops, the two came out level on software engineering, 92.4 percent against 92.6 percent. But K3 spent about 55 turns and 1.3 million tokens per task where Fable spent 21 turns and 130,000. Multiply that out at K3's own list price and the per-token discount stops covering the token count. A model that is a third of the price and ten times as verbose is not a third of the price.

RelatedKimi K3 Needed 10x the Tokens to Tie Fable 5 on SWE

That is the real reason the weights matter. If you are running enough volume that K3's turn count is the dominant cost, owning the deployment and controlling the context, the sampling and the stopping conditions is worth more than a headline per-token rate. That is a fleet-scale argument, and it is the population that can also afford eighteen accelerators.

  1. Jul 16K3 launches, API only 2.8T MoE, 1M context, $3/$15 per 1M tokens
  2. Jul 18vals.ai scores it independently 93.4% on SWE-bench Verified, bash-only harness
  3. Jul 22White House distillation accusation sanctions and Entity List designation floated
  4. Jul 27Weights deadline no repo on Hugging Face as of publication
  5. Aug 2026Third-party hosting and quantizations the point at which open weights change anyone's cost

Is it even open?

The expected license is Modified MIT, which is what Moonshot used for K2.6 and K2.7 Code. Expected is doing work in that sentence. No license text has been published for K3, and commercial usability is unconfirmed until it is. We flagged this on July 19 and it remains true: K3 sits on our AI coding leaderboard as a closed model, not an open one, because a roadmap commitment is not a downloadable file.

It earns its place there on merit regardless. vals.ai scores K3 at 93.4 percent on SWE-bench Verified using a neutral bash-only harness, fourth of 75 systems and the highest-scoring model on the board that anyone describes as open. Arena's blind developer voting puts it first on frontend code at 1,679 points, ahead of Fable 5.

What to watch · next 30 days
  • The license text, not the weights. Modified MIT with no field-of-use restriction makes this the largest genuinely permissive release ever. Anything narrower and the story changes completely.
  • Which format ships first. If Moonshot publishes MXFP4 rather than bf16, that is a deliberate choice to make the model deployable rather than maximally faithful, and it caps what anyone can fine-tune.
  • Serving price on third-party clouds. Below $3 per million input within a month would confirm the usual open-weight price collapse. Above it means the hardware floor is real.
  • Whether the distillation claim gets tested. Published weights are what would let independent researchers actually probe it. Watch for that work rather than for more statements.

Our take

Open weights at this scale are a genuine milestone and also a slightly hollow one. The word open has been carrying an implication since Llama 2 that a determined individual could get the thing running on hardware they could plausibly buy. At 2.8 trillion parameters that implication quietly breaks. What is being released is auditability and the right to build derivatives, which is real and valuable, but it is not access.

The honest framing is that K3's weights are a gift to cloud providers, researchers and distillers, not to developers. That is still worth having. It is just worth naming, because the gap between open-weight and open-in-practice is going to keep widening as frontier models get larger, and the vocabulary has not caught up.

And the deadline slipping past a morning is not the story. If Moonshot ships today, it ships. If it ships next week, the arithmetic is identical.

Primary sources

Original analysis by GenZTech. Memory figures are arithmetic from the published 2.8T parameter count. Hugging Face repository state checked directly on July 27, 2026. Hardware requirements via TECHi.