AI model router · 100+ models, one API

Don't choose a model.
Let MeshMoE pick
the best and cheapest one.

MeshMoE automatically selects the cheapest strongest model for you.
Not forwarding by price — it actually reads your question and dispatches the best model for the job:
coding questions to the best coder, math to the best reasoner. You just ask.

GPT-5.6-level quality at 1/10 the price.

what's in a name

Mesh + MoE. Not a brand. A thesis.

Modern frontier models — DeepSeek-V3, GLM-5.2, Qwen3 — are all Mixture-of-Experts. The intelligence doesn't live in one big brain. It lives in many small specialists, routed per task. We realized: that's exactly what the internet looks like too.

Mesh
/mɛʃ/ — a network of peers

The internet, treated as one big computer.

Millions of home PCs, gaming rigs, workstations — each with idle GPU time. Today that compute is wasted. Mesh means connecting them into a single addressable network, with no central server owning the capacity.

MoE
Mixture of Experts — the LLM architecture

Many specialists, one intelligence.

A router looks at each request and picks the right expert for the job. DeepSeek-V3 does this internally with 256 neural experts. MeshMoE does it externally — your request gets routed to the right node, the right model, the right tier of compute.

The isomorphism — same structure, two scales:

Inside a MoE model The MeshMoE network
Experts 256 small neural networks N home computers, each running a model
Router Learned gating network Our scheduler: reputation · load · latency · uptime
Activation 8 of 256 experts per token (sparse) 1 best-fit node per request (sparse)
Total capacity 671B params unbounded — grows with every new node
Owner One company no one — public infrastructure

This isn't a marketing metaphor. It's the same idea at a different scale. The architecture that lets a 37B-active model punch like a 671B one is the architecture that lets a network of small computers punch like a data center.

how it works

Every request, transparently routed.

Just pick mesh-pro (or any tier). MeshMoE reads your question and selects the best model automatically. Relay stations forward by price. We match by capability. That's the difference.

developer@api.meshmoe.com ~ curl
$ curl https://api.meshmoe.com/v1/chat/completions \ > -H "Authorization: Bearer moe-xxxx" \ > -d '{"model":"mesh-pro","messages":[{"role":"user","content":"write quicksort in Python"}]}' # MeshMoE reads this is a coding task, picks the best coder: X-MeshMoE-Capability: code X-MeshMoE-Routed: deepseek-v4-flash (coding 9/10) X-MeshMoE-Compressed: context -42% # Response body: { "model": "mesh-pro", "meshmoe_source": { "capability": "code", "routed_to": "deepseek-v4-flash", "compressed": "42% context saved" }, ... }

Relay station

  • Forwards by price — doesn't understand your task
  • Gives you 100 model names, you figure out which to use
  • Source hidden — you don't know what ran your prompt
  • You can only consume, never participate
  • Forwards at full price, no compression, no caching

MeshMoE

  • Reads your task, matches by capability — coding to the best coder
  • Four tiers, no model expertise needed (mesh-free / pro / max / research)
  • Fully transparent — shows who it picked and how much you saved
  • Anyone can contribute compute (phone or PC) and earn credits
  • Auto-compresses + caches — same quality, lower cost
network status · live

Honest numbers. Even when small.

Online experts
0 / growing
Edge hit rate
%
Total inferences
1 all-time
Open-weight models
41

// We don't fake traction. Today it's small. Be one of the first 100 nodes.

become an expert

Three commands. Your device joins the network.

The node client is open-source (MIT). Audit it before you run it — it should not mine, should not exfiltrate data, should only do what you tell it. Phones can contribute while charging. PCs earn credits in the background. Only real edge hits pay.

you@home-pc ~ bash
# 1. Download the open-source client (audit it first) $ curl -fsSL https://meshmoe.com/install.sh | bash # 2. Pick your expert (matches your hardware tier) $ meshmoe run --model qwen3-8b # 6GB+ VRAM · Tier 1 · credits ×1 $ meshmoe run --model qwen3-14b # 12GB+ VRAM · Tier 2 · credits ×3 # 3. That's it. Background process. Earn while you sleep. [meshmoe] registered as edge-node-7a3f (tier: 2, multiplier: ×3) [meshmoe] model loaded: qwen3-14b (Q4_K_M, 8.6 GB) [meshmoe] task #482: completed · +9 credits (300 tokens × ×3) [meshmoe] task #483: completed · +6 credits
// tier 0 · phone

Phone · contribute while charging

Gemma-2B / Qwen2.5-1.5B. Runs in background when charging + screen off. Handles simple tasks. Anyone can join. Credits ×0.3.

// tier 1 · light

Entry GPU · 6-12GB VRAM

Qwen3-8B / GLM-4-9B. RTX 4070 / 3060 Ti. General chat & quick tasks. Credits ×1.

RECOMMENDED
// tier 2 · standard ⭐

Gaming GPU · 12-24GB VRAM

Qwen3-14B / Kimi-K2. The main workhorse tier — highest routing rate. Credits ×3.

// tier 3 · heavy

Workstation / Large-memory · 24GB+

GLM-5 / Qwen3-32B / V4-Pro. RTX 5090 / A6000 / DGX Spark / M5 Ultra. Personal high-end devices. Deep reasoning + flagship models. Credits ×10.

// tier 4 · compute center

Compute Center · H100/A100 cluster

Enterprise-grade. Run full GLM-5.2 / DeepSeek-V4-Pro class flagships that no home PC can. Enterprise compute partners only. Credits ×30. Apply →

// revenue

Contribute compute, earn credits

When your node serves a request, you earn credits scaled by tier (×0.3 to ×30). Credits are spendable on mesh-pro/max/research, not withdrawable for cash. Only real edge-served work pays — no idle payout.

// future · personal expert

Fine-tune your own expert

Have private domain data? Future release supports local LoRA fine-tuning of personal expert models, contributed as specialized nodes. Read the vision →

// ready to become an expert?
Create your account →
or read the full setup guide →
research · not a product yet

The long dream: a GPT-class model, fragmented across homes.

Today, each node runs one whole small model — a "coarse-grained expert." The dream is finer-grained: what if a home computer could run one expert of a 671B model? 256 homes, 256 experts, one router — and the internet itself becomes a frontier AI.

This is hard. Petals tried layer-splitting (9K stars, "workable for batch jobs"). Nobody has shipped expert-splitting across WAN. We are honest about that. But the MeshMoE network — its router, scheduler, reputation layer, billing — is the natural testbed for whoever finally solves it.

If you're a researcher working on distributed MoE inference, cross-WAN KV cache sync, Byzantine-resilient inference, or dynamic router routing — talk to us. We'll provide the network. You provide the protocol.

// Personal Expert Models — your private data, your expert

A lawyer with 30 years of cases. A doctor with thousands of patient files. A researcher with internal notes never published online. GPT-4o will never have this knowledge — these data never went online.

MeshMoE's future client will let you fine-tune your own expert model (DeepSeek-R1-Distill + LoRA, trained locally — your data never leaves your machine). That expert joins the network as a specialized node. When others query your domain, the router sends them to you. Your private expertise earns credits while you sleep.

// This is the real Mesh of Experts — not a model architecture, but a network of human expertise, distilled into machines.

"The AI revolution will be distributed. Or it won't be a revolution at all."