📍 Home › Local AI

Project Maya: the Scheduler That Puts 748B on Your Desk

Local AI  Updated 2026-11-04 · 3 min read · #engine #moe #scheduler #exl3

What it is

Project Maya is a scheduler more than a launcher. Its engine profiles your GPUs, CPU, RAM and SSD, then splits each MoE model across three tiers: the experts your conversation uses most live on the GPU, the next ones in system RAM, the rest streams from NVMe — and they swap as you chat. One NVIDIA card (V100/RTX 20 or newer, up to 16 cards together); AMD and Windows are experimental, Mac is Metal-experimental, and it is explicitly not WSL2.

Getting started (3 steps)

1. git clone the repo, run ./setup.sh (Windows: START-MAYA.bat) — it asks which model you want, shows every download before starting it, builds once (20–40 min) and starts the model.
2. Open http://127.0.0.1:8080: browser chat, a live hardware monitor, a Models page, thinking levels, a one-click “Turbo” speed switch.
3. Point your tools at it through the OpenAI- and Anthropic-compatible APIs (Claude Code works with ANTHROPIC_BASE_URL=http://127.0.0.1:8080) — tool calls stream as the model writes them.

What it runs (sizes are the model’s weights; disk needs more)

Best use

When your workload is agentic: OpenAI/Anthropic APIs with tool calls, up to 8 requests answered at once (several agents, or several people, on one PC), a “saved vs. cloud-API-pricing” meter that shows what the same tokens would have cost at Z.ai’s or MiniMax’s rates, and drafters (DSpark/DFlash2) that raise decode speed on rigs where other paths stall.

Honest weaknesses

• It eats your RAM: the engine pins all but ~6 GB of free memory for its expert pool — other programs get the scraps while it runs.
• Fast NVMe is mandatory (~75–350 GB per model set); a slow SSD lowers every number above.
• Young project: experimental AMD/Windows paths, formats borrowed from llama.cpp and EXL3, and the popular DFlash2 drafter is CC BY-NC-ND — non-commercial only.

FAQ

Maya or Strata?

Roughly: Strata is the friendly launcher that ships the Qwen3.8-Flash-Next packs; Maya is the scheduler that runs the 320–748B trio above with a self-tuned expert pool. Several community packs run on both — when they do, Maya’s scheduling + draft-layer handling is what the crowd reports winning on the big models.

Why do people say its drafter beats ExLlamaV3?

On the big MoE models, community benchmarks showed plain EXL3 paths losing speed as context grew; Maya’s DSpark/DFlash2 draft layers went the other direction (e.g. 55–69 tok/s from ~31 on 4× R9700s). Treat every speed number here as hardware-specific — bench your own box with --bench.

Does anything leave my machine?

No — that’s the selling point. Chat history lives in your browser, the model runs in your box, and --report only writes a text file you can paste into a GitHub issue.