Kimi K3 is the strongest open-weight model (2.8T parameters, 896 experts) — but traditional deployment
needs 12-16× H200 GPUs. MeshMoE Cell splits K3 into its 92 layers and spreads them across
60-300 ordinary machines on your LAN.
Zero hardware purchase · data never leaves your network · earn 70% revenue share serving K3.
Your LAN is the AI datacenter you didn't know you had.
K3 weights are open (~1.4TB in MXFP4). "Open" doesn't mean "affordable to run":
| GPU cluster (vLLM / SGLang) | Cloud API | MeshMoE Cell (this) | |
|---|---|---|---|
| Hardware | Extreme 12-16× H200 (141GB) or 8-10× B200 — $400-600K+, plus InfiniBand and ops staff | None pay per token | Zero capex 60-300 PCs you already have (32GB RAM each) |
| Data privacy | Full on-premise | Leaves prompts go to a third-party cloud | Full all compute inside your LAN; any single machine only sees meaningless intermediate vectors |
| Cost model | $400K+ upfront, or $35-70K/month rented | $3/M input · $15/M output — adds up fast | ≈ electricity only (machines are on anyway) |
| Throughput / speed | ~3,000 tok/s · 30-50 tok/s per user | provider-limited | ~500-850 tok/s per 200-machine Cell · 4-7 tok/s per user (fine for research, coding, agents, batch) |
| Revenue? | pure cost center | — | Yes: serve K3 to MeshMoE's global users, keep 70% |
| Best for | Big tech, national labs | Low-volume teams | Internet cafés · esports hotels · school labs · office parks · small IDCs |
Kimi K3 is a stack of 92 nearly identical MoE layers. Each layer — all 896 experts of it — is just 14.8GB in MXFP4. So instead of one impossibly large server, each machine holds exactly one layer, and each request flows through 92 machines like an assembly line. A full round trip produces K3's answer.
| You might ask | Answer |
|---|---|
| Does every PC download the whole model? | No. One layer each — 15.4GB, downloaded once over your LAN from peer machines. Smaller than a AAA game. |
| Will it eat our internet bandwidth? | No. The 92-hop pipeline runs entirely on your LAN switch (7KB per hop). WAN traffic is just lightweight heartbeats and task dispatch. |
| Will machines be pegged at 100%? | No. One K3 layer is ~66 MFLOPs per token — microseconds of work. Machines idle between batches; customers can game on the same PC without noticing. |
| Who orchestrates 92 machines? | We do. MeshMoE's scheduler assigns layers, builds batches, and reroutes around failures automatically. You just keep the machines on. |
Search "kimi k3 vram requirements" and you'll find despair: the full model needs 1.4TB of VRAM. Cell changes the equation — each machine carries 1/92 of the model. Mixed fleets are fine; our scheduler assigns roles by detected hardware.
| Machine spec | Eligible? | Role |
|---|---|---|
| 32GB RAM + gigabit NIC standard in post-2023 cafés / offices |
✅ Standard node | holds one full K3 layer (15.4GB) in RAM |
| VRAM ≥16GB (4070 Ti S / 4080 / 4090) | ✅ Premium node | layer lives in VRAM — fastest; prioritized for attention-heavy layers |
| 16GB RAM + 8GB VRAM | ⚠️ Paired mode | two machines co-host one layer (7.7GB each), auto-paired |
| 16GB RAM, no local disk | ✗ Not for K3 | can still join MeshMoE's classic edge-node program (see homepage) |
| Other requirements | Details |
|---|---|
| Disk | ≥20GB free on a local SSD. Diskless (PXE-boot) systems need one small local SSD, or layer shards can live on your in-house file server — we survey your setup before signing. |
| Network | Gigabit LAN minimum, 10GbE nice-to-have. No special WAN requirements. |
| OS | Windows 10/11 with the MeshMoE desktop client (batch installer provided). |
| Scale | 60 machines minimum · 100 stable · 200 ideal (two redundant K3 pipelines). |
One rule of thumb: 60+ powered-on machines on the same LAN. More machines = faster + richer.
Post-2023 venues (32GB RAM machines) are all-standard-node ready. Mixed old/new fleets work too.
Uniform hardware and central IT make this the smoothest deployment of all. Scheduled on/off supported.
Data stays inside your corporate LAN. Auto-pause during work hours.
Turn spare racks into a 24/7 K3 service node — better yield than plain colocation.
Your Cell serves K3 at half the official API price ($7.5/M output tokens). You keep 70%; the platform (30%) brings the customers, orchestration, cloud fallback and billing:
| Cell utilization | Gross / month | Your 70% | Electricity | Net / month |
|---|---|---|---|---|
| 25% (conservative, ramp-up) | ~$4,700 | ~$3,300 | ~$1,250 | ~$2,000 |
| 50% (healthy, steady) | ~$9,400 | ~$6,600 | ~$1,250 | ~$5,300 |
| 100% (theoretical max) | ~$18,800 | ~$13,100 | ~$1,250 | ~$11,900 |
On electricity: by day your machines are on for customers anyway — Cell adds ~10% on top. Overnight batch mode costs roughly $8/night for 200 machines. Figures above assume always-on plus summer cooling overhead — deliberately conservative.
The full model is ~1.4TB of weights (MXFP4). Official cluster deployment needs 12-16× H200 (141GB each) — a single 4090 (24GB) cannot hold K3, period. But one 4090 holds one Cell layer (14.8GB) comfortably and serves as a premium node. So: one 4090 can't run K3 — 92 ordinary machines together can.
vLLM/SGLang assume the full 1.4TB model fits in a few monster GPUs — $400K+ of hardware plus ops staff. Cell distributes K3 layer-by-layer across commodity PCs you already own: zero capex. The tradeoff is per-user speed (4-7 tok/s vs 30+), which is why Cell targets research, coding, agent and batch workloads rather than snappy chat.
Official API: $3/M input, $15/M output tokens. If your organization burns more than ~200M tokens/day, a Cell you own is an order of magnitude cheaper — and it can earn money back by serving external traffic at night.
Yes — released by Moonshot AI on 2026-07-28 under the Kimi K3 License, on Hugging Face (moonshotai/Kimi-K3) and ModelScope. But it's ~1.4TB and won't run on one machine. Cell partners don't download the model at all: each machine pulls its own 15.4GB layer from LAN peers. We operate strictly within the license terms.
Barely. Each machine computes one layer — milliseconds of work per batch — leaving CPU/GPU nearly idle on average, plus ~16GB RAM. Customers game on the same PCs without a visible FPS change. Offices can set work-hours auto-pause.
Every machine only sees one layer's intermediate activations — meaningless number vectors. No single machine can reconstruct any user's prompt; that's architectural privacy, not policy. Platform-side data compliance is contractually ours, not yours.
The scheduler detects a dead node in ~5s, shifts traffic to the backup pipeline, and a hot spare takes over within 30s. Your earnings accrue by actual uptime — no penalties for machine failures.
Organizations: monthly B2B settlement with invoices. Individual node operators (future metro Cells): you earn MeshMoE credits proportional to your contribution, which offset your own AI usage on the platform (credits are not redeemable for cash). No crypto, no tokens, ever.
Pilot partners get priority traffic to saturate their Cell plus a 3-month electricity floor subsidy. What we need: ≥60 powered-on machines on one LAN, one cooperative admin, and your curiosity.
Reply within 1 business day · free remote site survey · live K3 deployment within 2 weeks of signing