Kimi K3 Open Weights: What It Costs to Self-Host
The Short Answer
Kimi K3’s full weights release by July 27, 2026 under a Modified MIT license — but “open weights” does not mean “runs on your laptop.” At 2.8 trillion parameters (Mixture-of-Experts), Kimi K3 is the first open model in the 3T-parameter class, and self-hosting it is a data-center project, not a weekend one.
The model has been available on the Kimi API since July 16, 2026 at $3 input / $15 output per MTok (flat pricing, no peak/valley surge). The July 27 weights drop is about control — inspection, fine-tuning, air-gapped deployment — not about saving money.
What Kimi K3 Actually Is (July 2026)
| Spec | Kimi K3 |
|---|---|
| Developer | Moonshot AI |
| Total parameters | 2.8 trillion (MoE) |
| Context window | 1M tokens |
| Architecture | Kimi Delta Attention (KDA) + Attention Residuals |
| Modality | Text + native visual understanding |
| API pricing | $3 / $15 per MTok (flat) |
| API launch | July 16, 2026 |
| Weights release | By July 27, 2026 |
| License | Modified MIT |
| Artificial Analysis Index | ~57 |
Kimi K3 leads the Arena.ai Frontend Code Arena ahead of Claude Fable 5 and sits near the top of the open-weight tier on the Artificial Analysis Intelligence Index (~57).
The Hardware Math
The single most important spec for self-hosting is VRAM — the active experts plus KV cache must fit in GPU memory for usable speed.
| Tier | Config | What runs |
|---|---|---|
| Frontier (native) | 16x H100 80GB (~1,280GB) or multi-B200 | Kimi K3 at full quality |
| Quantized | 8x H200 / multi-node H100, INT4/MXFP4 | Kimi K3 with quality loss |
| Workstation | 1x RTX 4090 (24GB) | NOT Kimi K3 — use GLM-5.2 or DeepSeek V4 Flash |
Reality check: even at INT4/MXFP4 quantization, a 2.8T MoE model’s weight footprint is enormous. If your goal is “open weights I can run on a single 24GB card,” Kimi K3 is the wrong model. GLM-5.2 (744B, MIT) or DeepSeek V4 Flash are the practical self-host picks in July 2026.
Self-Host vs API: The Cost Decision
Use the API ($3/$15 per MTok) when:
- Your volume is low or spiky
- You can’t keep a 16x-H100 cluster near full utilization
- You want zero infra maintenance
Self-host the weights when:
- You need air-gapped / data-residency compliance
- You’re fine-tuning or modifying the model
- You have sustained, high-volume, predictable throughput to amortize the cluster
At low-to-medium volume, idle GPU time makes self-hosting more expensive than the API — the classic open-weight trap. The weights give you control; they rarely give you cheaper inference below cluster-saturating scale.
What to Do This Week
- If you just want to use K3: the API at $3/$15 is live now — no need to wait for weights.
- If you want open weights to run yourself: benchmark GLM-5.2 first; it’s self-hostable today, MIT-licensed, and ~1/3 the output cost of K3.
- If you specifically need K3 self-hosted: budget for a multi-node H100/H200 cluster and plan for INT4/MXFP4 quantization from day one.
- Watch the license text — “Modified MIT” may carry deployment or attribution conditions; read it before commercial use.
Sources
- Kimi K3 quickstart & specs (Moonshot AI): platform.kimi.ai/docs/guide/kimi-k3-quickstart
- Kimi K3 self-hosting overview (Northflank, July 2026): northflank.com/blog/what-is-kimi-k3-self-hosting
- Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2 serving cost (MarkTechPost, July 18, 2026): marktechpost.com
- Moonshot AI: moonshot.ai