Kimi K3's Weights Are Free — the 1,561 GB You Need to Hold Them Are Not

Moonshot AI published the full weights for Kimi K3 on 27 July, in 96 files totalling just over 1.5 terabytes. The model itself launched on 16 July. This release is the parameters — the part that matters if you want to run it yourself.

The launch itself, and the specifications Moonshot published with it, are covered in our earlier report: Moonshot releases Kimi K3, the largest open-weight model yet, with weights due 27 July.

Most coverage will stop at "2.8 trillion parameters, open weights". We downloaded the file manifest and did the arithmetic on what that actually means — including one check that the model card cannot do for itself.

What was published

The headline number is 2.8 trillion parameters, but only 104 billion are active on any given token — 3.71% of the model. Sixteen of 896 experts are selected per token, or 1.79%. The result is a memory-to-compute ratio of 26.9 to one. You pay to store a 2.8T model but get the inference cost of a 104B one.

That trade is the entire design, and it is also why the storage figure below is the binding constraint rather than the FLOPs.

The file sizes check the model card

The model card states the weights are MXFP4 — a four-bit format with a shared scale factor per 32-value block, which works out to 4.25 bits per parameter. That is a claim about the file you are downloading, so the file can be used to test it.

Summing all 96 shards gives 1,560,936,091,448 bytes. Divided by 2.8 trillion parameters, that is 4.460 bits per parameter — 0.210 bits above pure MXFP4, or roughly 73.4 GB of the model held at higher precision.

That residual is expected. Embeddings, normalisation and routing layers are routinely kept at higher precision, since quantising them costs disproportionate quality. The small excess is not the point; the published bytes being consistent with the published architecture is. An independent check that agrees is worth more than a claim on its own.

Free to licence, expensive to hold

Twenty H100s are needed just to hold the weights. That figure excludes the KV cache, activations and any headroom for concurrent requests, so a real deployment needs materially more than the floor in that table.

Then there is getting it. At 1 Gbps — a common Singapore business fibre line — 1,560.9 GB takes about 3.47 hours to transfer. On a 100 Mbps connection it is 34.7 hours. On 10 Gbps, 21 minutes. If you want to convert those figures for your own link, our data storage converter handles the unit arithmetic.

None of this makes the release less meaningful. It does mean the practical audience for the weights is inference providers, well-funded labs and national compute programmes — not the individual developers for whom "open weights" usually reads as an invitation.

The price of not running it yourself

The API pricing cuts against the expected narrative. Moonshot's own rates for K3 sit in the premium tier.

At a 3:1 input-to-output blend, K3 costs US$6.00 per million tokens — above Gemini 3.1 Pro at $4.50, and eleven times DeepSeek V4-Pro at $0.54. It undercuts Claude Opus 5 by 40%, which is the comparison Moonshot would presumably prefer, but it is not the cheap-Chinese-model story the "open weights" framing invites.

One number does move sharply: cached input is US$0.30 per million against $3.00 on a cache miss, a tenfold gap. That saves money on repeated prompt prefixes, but it is a bad planning assumption for anything else. Every comparison above uses cache-miss pricing for that reason. If you want the same arithmetic across the Chinese labs specifically, we keep a Chinese LLM comparison current.

The licence has more teeth than "open" suggests

The permission grant is broad — use, copy, modify, distribute, sublicense, sell, fine-tune. Two conditions narrow it.

Clause 2 is the one to read twice. If you operate a Model-as-a-Service business and the aggregate revenue of your company and its affiliates exceeds US$20 million over any consecutive twelve months, you must enter a separate agreement with Moonshot before any commercial use. Note that the threshold measures total company revenue, not revenue generated by the model. A mid-sized software company offering fine-tuning access crosses it on the strength of its existing business, having earned nothing from K3 at all.

Clause 3 is lighter — a badging requirement at 100 million monthly active users or $20 million in monthly revenue. Clause 4 exempts internal use entirely, which is where most enterprise deployments will sit, and use through Moonshot's own products or certified partners.

The caveats that matter

• We have not run the model. Nothing here speaks to output quality. This is an analysis of what was published — file sizes, architecture, prices and licence terms — not an evaluation.
• The bits-per-parameter figure assumes the stated parameter count. We divided measured bytes by Moonshot's own 2.8T figure; we did not independently count parameters from the tensor metadata.
• Hardware counts are a floor, not a configuration. Dividing weights by accelerator memory ignores KV cache, activation memory, tensor-parallel overhead and the fact that you cannot use 100% of a card's memory.
• Download times are theoretical maxima at the stated line rate, with no allowance for real throughput, contention or per-connection limits.
• Benchmark claims are excluded deliberately. The model card reports scores across a range of benchmarks; those are self-reported and we have not verified them, so we have not built a table on them.
• Prices exclude tax, per Moonshot's own note, and rival rates were verified on 25 July — API pricing moves quickly.

Key takeaways

• What shipped. Full Kimi K3 weights, 27 July 2026: 96 shards, 1,560.9 GB, 2.8T parameters with 104B active.
• The card checks out. 4.460 measured bits per parameter against 4.25 for pure MXFP4 — consistent with a four-bit model keeping some layers at higher precision.
• The floor to hold it is 20 H100s, 12 H200s or 9 B200s — 11.3% of a GB200 NVL72 rack — before any serving overhead.
• The API is not cheap. $6.00 per million blended 3:1: above Gemini 3.1 Pro, 11× DeepSeek V4-Pro, 40% below Claude Opus 5.
• The licence requires a negotiation from any MaaS operator whose total revenue exceeds $20m over twelve months — a threshold about company size, not model usage.