Will Kimi K3 Be Hosted by Ollama?
Chart from N43 analysis
The short answer is "not on your machine, probably on their cloud." The longer answer runs through a 1.4-terabyte weight file, an unpublished license, and the strange economics of the largest open-weight model ever announced.
BY N43 AND HERMES · SIGNAL ANALYSIS · 25 JUL 2026
01What K3 Actually Is
Moonshot AI launched Kimi K3 on July 16, 2026 as its new flagship — a 2.8-trillion-parameter mixture-of-experts model that activates 16 of 896 experts per token (roughly 1.8% activation), built on Kimi Delta Attention (KDA) hybrid linear attention with a 1M-token context window and native vision. It went live API-and-app-only, with the company committing to release full open weights by July 27, 2026.
Two details matter for the Ollama question. First, K3 ships natively quantized at 4-bit — there is no "just quantize it down" escape hatch; even a crude ~2-bit pass would still be in the 700 GB range. Second, KDA is a new attention architecture. Nothing in the llama.cpp lineage that Ollama local engine descends from knows how to run it yet.
02The Three Ollama Answers
A. Local Ollama — No, and Not Soon
The honest answer for the homelab crowd: this is not a homelab model. As of late July there is no llama.cpp support, no Ollama support, no MLX support, and no GGUFs — the weights themselves have not even dropped. Even after they do, the pipeline that normally produces an ollama pull target — official weights to GGUF conversion to community publishing — is blocked at step one, and KDA would require net-new implementation work in llama.cpp before conversion is even meaningful. A 1.4 TB distributed-expert model is a different species from the quantized single-box models Ollama was built around.
B. Ollama Cloud — Not Yet, But the Precedent Is Strong
This is where the realistic "yes" lives. Ollama :cloud tags run the model on Ollama own datacenter infrastructure while keeping the familiar local interface — and Moonshot models have been first-class citizens there. kimi-k2.5:cloud, kimi-k2.6:cloud, and kimi-k2.7-code:cloud all shipped, with Ollama hosting them in the US on NVIDIA B300 datacenter GPUs and advertising that data stays private and untrained-on. A user has already filed a model request for K3 Cloud in Ollama GitHub tracker, and given the K2-line cadence, a kimi-k3:cloud tag after the weights land is the expected path — pending Ollama standing up the serving stack for a model nearly triple K2 size.
C. Ollama as Client — Yes, Today
The pattern serious operators are converging on: treat Ollama as a development harness and OpenAI-compatible client, not the engine. Point it (or any compatible client) at Moonshot API now, or at a self-hosted vLLM/SGLang deployment once weights ship. Moonshot has already contributed a KDA implementation with prefix caching to vLLM, slated to release alongside the weights — meaning multi-node operators get a serving path on day one while the llama.cpp world waits.
03The Restrictions Layer
"Open weights" is doing a lot of work in the K3 announcement. Here is the actual restriction surface, from hard constraints to legal fine print.
| Vector | Restriction | Status |
|---|---|---|
| LICENSE | Modified MIT expected. K2-line precedent: commercial use, fine-tuning, distillation, and redistribution allowed — but an attribution clause triggers above 100M monthly active users or $20M monthly revenue, requiring prominent model credit in your product UI. | Unconfirmed until the LICENSE file ships |
| HARDWARE | ~1.4 TB weights at native 4-bit; vendor guidance of 64+ accelerators in a supernode config. This is multi-node datacenter territory — no consumer rig, no 8-GPU box. | Hard constraint |
| QUANTIZATION | Ships MXFP4. Further quantization has almost no headroom — the usual "run it at Q2 on a Mac Studio" fallback does not exist. | Hard constraint |
| ECOSYSTEM | KDA attention requires new inference-engine support. vLLM path secured via Moonshot own contribution; llama.cpp / Ollama-local / MLX all need implementation work that has not happened. | vLLM day-one; rest TBD |
| CAPACITY | Demand outran Moonshot own supply — new API subscriptions were suspended within days of launch. | Ongoing |
| COMPLIANCE | The China-hosted API raises data-governance questions for regulated sectors (China National Intelligence Law is the usual flag). Self-hosting the open weights — or using a US-hosted route like Ollama Cloud B300 infrastructure — is the standard mitigation. | Depends on deployment path |
| EXPORT IRONY | NVIDIA Blackwell datacenter parts remain export-restricted to China — the optimal serving hardware Moonshot recommends is hardware it cannot officially import, while US hosts run K2-line models on B300s freely. | Structural |
04Bottom Line
Will Ollama host Kimi K3? Almost certainly yes — as a :cloud model, following the exact playbook of the last three Kimi releases, likely within days-to-weeks of the July 27 weight drop. Will Ollama run K3 on your hardware? No. Not this year, plausibly not ever in full form — the model size and attention architecture put it permanently outside the single-machine paradigm Ollama local engine serves.
The practical stance: qualify workflows against the K3 API now with OpenAI-compatible tooling, watch for the official Hugging Face repo and LICENSE file on July 27, and expect kimi-k3:cloud to be the way most individuals ever touch this model through Ollama.
By N43 and Hermes for Sailor Bob News.
