Which Software Runs the Big Three? Strata, Maya, and the GGUF World, Compared

Home Rig Updated 2026-11-04 · 3 min read · #engine #software #strata #maya #gguf

Status board (checked 2026-11-04)


Strata	guided launcher, Windows + Linux; NVIDIA, AMD, Intel Arc, Apple Silicon paths; home of the Flash-Next packs
Project Maya	three-tier scheduler (GPU → RAM → NVMe), MIT; NVIDIA V100/RTX 20+ up to 16 cards; OpenAI- and Anthropic-compatible APIs
llama.cpp / LM Studio / Ollama	the GGUF standard — widest model coverage, community quants (Unsloth, Vultr, AtomicChat)
Licenses to note	Maya’s popular DFlash2 drafter is CC BY-NC-ND — non-commercial only

What it is

Same models, three philosophies:

Strata plans before it downloads: strata launch --plan prints what fits your machine, then installs exactly that. Its --tune quick profiles your box and draft layers (MTP ≈6 GB) come automatically.
Maya schedules while you chat: hot experts on the GPU, next ones in RAM, rest on NVMe — moved around as your conversation drifts. Up to 8 concurrent requests, a “saved vs cloud-pricing” meter, tool calls streaming through OpenAI/Anthropic endpoints.
The GGUF world (llama.cpp, LM Studio, Ollama) runs everything else — and the community quant scene (imatrix, hybrid per-layer precision) keeps producing variants the engines don’t ship.

Getting started (3 steps)

Start from your RAM, not the hype: pick the pack size, then let Strata’s plan or Maya’s setup recommend a tier — both show every download before it starts.
Agentic work → Maya’s endpoints: point Claude Code / OpenCode / any OpenAI SDK at http://127.0.0.1:8080 (Maya documents the ANTHROPIC_BASE_URL one-liner) and tool calls stream as the model writes them.
Bench your own box (--bench, --tune quick) before trusting any viral speed clip — every number on our model pages is hardware-specific.

Best use

You want…	Pick
The friendliest first run of Flash-Next	Strata — the plan shows what fits before anything downloads
GLM / MiniMax / DeepSeek at home, several agents at once	Project Maya — the scheduler + dual APIs are the point
Any other model, widest compatibility, Apple Silicon	GGUF tooling — check quantization trade-offs first

Honest weaknesses

Quant packs are not interchangeable: Maya packs, GGUF, EXL3 and NVFP4 are different formats with different sizes for the “same” model — check the quantization page before mixing.
Maya pins nearly all your RAM while running (other programs get the scraps) and is NVIDIA-first; AMD and Windows paths are experimental.
Strata’s best numbers live in its docs, not one-click defaults; answers can shift between runs as experts move.

FAQ

Strata or Maya — one line each?

Strata = the guided launcher for the Flash-Next family. Maya = the scheduler for the 320–748B trio with agent APIs. Community packs often run on both.

Do I still need LM Studio?

For browsing GGUF of the thousands of models outside the Big Three — yes, it’s the nicest catalog. For the Big Three, the engine-specific packs are where the measured speeds live.

Can my phone or another PC use the home rig?

Maya serves OpenAI-compatible endpoints on your LAN by design; Strata runs local UIs. Either way nothing leaves your machine — that’s the selling point over renting.

📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.

← Newer: What Quantization Costs You: the IQ1→Q8 Ladder With Measured Numbers