Kimi K3's Open Weights Finally Landed — But Self-Hosting the Largest AI Model Ever Costs $3–4 Million in Hardware

On July 27, 2026, Moonshot AI published the full open weights for Kimi K3 on Hugging Face — 2.8 trillion parameters, the largest open-weight model ever released, under a Modified MIT license. Nine days earlier, this blog covered K3 topping coding-preference benchmarks against Claude Fable 5 and GPT-5.6 Sol while noting one catch: nobody could actually download it until today. That day has arrived, and the release shows both how far open-weight models have closed the gap with closed frontier labs on coding capability, and why 'open' no longer means 'easy to run' — self-hosting K3 requires roughly 1.4 terabytes of weights and a GPU rack that costs $3 to $4 million to buy.

A 2.8-Trillion-Parameter Model, 1.8% Active

Kimi K3 uses a Stable LatentMoE architecture with 896 experts, routing each token through just 16 of them — about 50 billion active parameters per step, or roughly 1.8% of the model's total weights. Two attention innovations, Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), are built to keep long sequences and deep layers stable, backing a 1,048,576-token (roughly 1 million) context window. Moonshot trained the model with quantization-aware training from the supervised fine-tuning stage onward, using native MXFP4 weights with MXFP8 activations, which is what lets the full model ship in roughly 1.4TB rather than the multiple petabytes a naive FP16 release of this size would require. The weights are published at huggingface.co/moonshotai/Kimi-K3 — nearly three times the size of K2's 1 trillion parameters, and the largest open-weight release to date.

How It Stacks Up Against Claude and GPT

On Terminal-Bench 2.1, Moonshot reports K3 at 88.3, just behind GPT-5.6 Sol's 88.8 and ahead of both Claude Fable 5 and Opus 4.8 at 84.6. SWE-bench Verified comes in at 76.8%, and K3 ranks #1 on Moonshot's own Program Bench (77.8) and SWE Marathon (42.0) suites. On the independent Artificial Analysis Intelligence Index, K3 scores 57.11 — #4 overall, behind Fable 5 (59.86) and GPT-5.6 Sol max (58.89), but ahead of Opus 4.8 (55.69). On the community-run Frontend Code Arena, K3 takes the top spot outright at 1,679 Elo, ahead of Fable 5's 1,631 and GPT-5.6 Sol's 1,618. None of that makes K3 the strongest model available — it trails on general intelligence — but it is now the strongest open-weight model by a wide margin, and competitive with closed frontier models on the coding tasks most developers actually care about.

The Catch: Open Doesn't Mean Affordable

Publishing the weights is not the same as making the model easy to run. Serving K3 yourself takes something like a 72-GPU NVIDIA GB200 NVL72 rack, which costs $3 to $4 million to buy and install, draws roughly 120 kW continuously, and burns over a million kilowatt-hours a year before counting networking, storage, cooling, or staff — renting equivalent capacity on the open market runs close to $7 million annually. The economics are almost the inverse of what you'd expect: because only about 50 billion of the 2.8 trillion parameters fire on any given token, K3 is comparatively cheap to run once it's loaded, but expensive to host because all 2.8 trillion still have to sit in fast memory at once. That's the whole tension in this release: the license is open, but the hardware bill is not.

Why Enterprises Will Self-Host Anyway

For most individual developers and small teams, the practical way to use K3 is Moonshot's hosted API — priced at $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens, with Moonshot reporting over 90% cache-hit rates on typical coding workloads. But the open weights matter most to a different audience: enterprises and government-adjacent teams wary of routing proprietary code through a China-based API, who can now self-host K3 on their own infrastructure to keep data resident and auditable — the same data-sovereignty logic that has shaped how regulated industries approach DeepSeek's releases. The Modified MIT license permits commercial use and fine-tuning, but Moonshot had not published K3's final license text as of the July 27 release; every prior K2-family release carried a monthly-active-user clause on large commercial deployments, so teams should read K3's actual license file rather than assume K2's terms carried over.

What It Means for Developers and AI Teams

For most teams, nothing changes operationally today — you can already reach K3 through Moonshot's API or Kimi Code, or a prior K2-family model through GitHub Copilot's model picker, without touching a GPU. What changes is leverage: a near-frontier coding model is now formally available to fine-tune, audit, and run entirely outside any vendor's servers, which gives procurement and security teams a credible alternative to point to when negotiating with closed-model vendors, even if few of them will actually stand up the hardware to use it directly. Cloud providers and sovereign-AI initiatives with existing large GPU fleets are the ones most likely to host it themselves; everyone else gets the option value without the capital expense.

Bottom Line

Kimi K3's release confirms that open-weight models have closed most of the coding-capability gap with closed frontier labs — on several benchmarks that matter to developers, K3 now beats Claude Fable 5 and GPT-5.6 Sol outright. What it also confirms is that the real barrier to running a frontier-class model yourself stopped being the license a while ago; it's now the $3-4 million hardware bill. That reframes the competition among frontier labs: it's no longer just a fight over who publishes the best weights, it's a fight over who makes running those weights affordable — and on that measure, hosted APIs, not open licenses, are still deciding who actually gets to use frontier AI.