Inside The Machine

Authored by Neal Lloyd · Daily AI Series

Inside The Machine

Issue 41 · AI Corner · Inside The Machine

Day 41

Open Weights · Mixture-of-Experts · Free vs. Usable

Kimi K3's Weights Are Free Today. Almost Nobody Can Actually Run Them.
Moonshot AI's 2.8 Trillion Parameter Model Just Became the Largest Open-Source-Track Release Ever. Here Is What “Open Weights” Actually Means at This Scale, and the Real Gap Between Legally Free and Practically Usable.

Ground Truth, Episode 31, covered the governance and pricing-pressure story behind Kimi K3's release: a model launched to upstage a rival, with full weights now public. This series takes the technical question the headline number obscures: what does open weights actually mean for a 2.8 trillion parameter model, and who can realistically do anything with them once they are downloadable? Day 41 explains the gap between legally free and practically usable.

Neal Lloyd

Author · Inside The Machine · July 2026

11 min read

"Open weights" sounds like a binary: either a model's parameters are downloadable or they are not. In practice it is closer to a spectrum, and at 2.8 trillion parameters, the distance between legally downloadable and actually usable by anyone without a dedicated GPU cluster is enormous. Kimi K3's weights being free today changes who is allowed to run it. It does not change who actually can.

Neal Lloyd · Inside The Machine, Day 41

Ground Truth Episode 31 covered the governance and pricing story behind Kimi K3's release: a deliberately timed launch, full weights now public, real competitive pressure on closed labs. This series takes the harder technical question underneath the headline number. What does 2.8 trillion parameters and open weights actually mean, mechanically, for who can use this model and how? This is Day 41 of Inside The Machine.

Section I — What "Open Weights" Actually Promises

Legally Free to Download Is Not the Same as Practically Usable

"Open weights" specifically means the trained parameters of a model are published for anyone to download, distinct from "open source," which would also require releasing the training code, data, and methodology. Kimi K3 falls into the open-weights category: the numbers that define the model's behaviour are public, but Moonshot's exact training process, data mixture, and infrastructure remain proprietary. This is the norm for essentially every major open-weight release, including DeepSeek's and Meta's Llama family.

At 2.8 trillion total parameters, even in a sparse mixture-of-experts architecture where only a fraction of parameters activate for any given request, the raw storage footprint alone runs into the terabyte range at typical precision. Downloading the weights is realistic for anyone with a fast internet connection and patience. Actually loading the model into memory and running inference at any usable speed requires hardware most individual developers, and a meaningful share of small companies, simply do not own.

This gap is not unique to K3 — it exists for every large open-weight release, and it is precisely why the phrase "democratising AI" attached to open-weight launches deserves scrutiny rather than automatic acceptance. Democratising access to the licence to run something is a real and valuable thing. It is not the same as democratising the practical ability to run it, which remains gated by hardware cost regardless of what the licence says.

⚡ Open Weights vs. Open Source, Plainly

Open weights: the trained model parameters are publicly downloadable. Open source: weights, training code, data, and methodology are all published, allowing full reproduction. Kimi K3, like DeepSeek and Llama before it, is open-weight, not full open-source — Moonshot's training process and data mixture remain proprietary even as the resulting model is free to download and run.

Section II — Why Sparse Mixture-of-Experts Changes the Math

2.8 Trillion Parameters, But Not All of Them Fire at Once

A dense model activates every one of its parameters for every request — a straightforward but computationally expensive design. A sparse mixture-of-experts model, the architecture K3 uses, instead divides its parameters into specialised sub-networks, or experts, and a routing mechanism selects only a small subset of them to activate for any given input. The 2.8 trillion figure describes K3's total parameter count across all experts combined; the actual compute cost of processing a single request is governed by how many parameters activate per token, which is a much smaller number.

This is precisely why a 2.8 trillion parameter sparse model can be more practical to run than a much smaller dense model at the same effective inference cost — the total size determines storage and memory requirements, while the active parameter count determines compute cost per request. It also means the hardware requirement question has two separate answers: how much memory do you need to hold the full model (governed by the 2.8 trillion figure), and how much compute do you need to actually process a request (governed by the much smaller active count).

The memory requirement is the harder constraint in practice. Even with sparse activation reducing compute cost, the full 2.8 trillion parameters generally need to be loaded into fast memory somewhere to make routing between experts efficient, which is why self-hosting a model at this scale still typically requires either a substantial multi-GPU cluster or specialised inference infrastructure — not a single consumer graphics card, regardless of how sparse the activation pattern is.

Sparse mixture-of-experts is a genuine engineering breakthrough for making enormous models computationally tractable to run. It is not a breakthrough for making them fit on a laptop. Those are two different problems, and only one of them got solved.

Neal Lloyd · Inside The Machine, Day 41

Section III — Who Actually Benefits From Today's Release

Cloud Providers and Well-Funded Labs, Mostly. That Is Still a Real Outcome.

In practice, the immediate beneficiaries of today's weight release are cloud infrastructure providers, well-capitalised AI labs, and research institutions with existing GPU clusters — the organisations that already had the hardware to run something at this scale and now have one more model to choose from at zero licensing cost. This is a real and meaningful outcome. It lowers costs and increases choice for a segment of the market that already had meaningful market power, even if it does not hand a 2.8 trillion parameter model to an individual developer's laptop.

The more diffuse benefit flows through API access. Moonshot and third-party providers will host K3 and sell inference access by the token, meaning developers without their own hardware still gain a new, cheaper option through the same hosted-API model this series has covered throughout the Chinese open-weight story — the model becomes accessible in practice even to people who will never personally handle the weights file. That is a genuinely different, and genuinely more consequential, form of access than the raw open-weight download represents on its own.

The honest framing, then, is neither "this democratises AI" nor "this changes nothing for ordinary developers." It expands the field of organisations that can host and compete on this model's capability, which indirectly benefits everyone downstream through cheaper API pricing and more competitive choice — the same dynamic this series has tracked since Chinese open-weight models first started reshaping enterprise routing decisions back in Episode 24.

Open weights rarely put a frontier model directly into an individual's hands. What they reliably do is add a new competitor to the small set of organisations that can host and sell access to something this capable — and that competition is where the real benefit to everyone else eventually shows up, in lower API prices, not in a download anyone's laptop can actually run.

Neal Lloyd · Inside The Machine, Day 41

— Neal Lloyd
Inside The Machine, Day 41 · July 28 2026

← PreviousDay 40: The DeepSeek Migration Deadline Nobody Is Talking About

Next →Coming Soon

About The Author

Neal Lloyd

Author · Series Creator

Authored by Neal Lloyd

Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.

By The Numbers

2.8T

Total parameters in Kimi K3, the largest open-source-track model release to date, with full weights available today.

1TB+

Approximate storage footprint of K3's weights at typical precision — downloadable by anyone, but requiring serious infrastructure to actually load and run.

Separate hardware constraints governing usability: total memory needed to hold all parameters, and compute needed per request based on active parameters only.

The Series

Key Concepts

Open Weights

A model release format where the trained parameters are published for download, distinct from open source, which would also require training code, data, and methodology.

Sparse Mixture-of-Experts

An architecture dividing parameters into specialised sub-networks (experts), with a routing mechanism activating only a small subset per request, reducing compute cost relative to a dense model of the same total size.

Active Parameters

The subset of a sparse model's total parameters that actually activate for a given request, governing per-request compute cost independently of the model's full size.

Legally Free vs. Practically Usable

This series’ distinction between a model being licensed for free download and an individual or organisation actually having the hardware to run it.

Hosted-API Access

The channel through which most developers actually gain practical access to a large open-weight model — via a hosting provider's API, rather than running the weights themselves.

Inside The Machine

An ongoing daily editorial series on artificial intelligence.

Authored by

Neal Lloyd