Strata and Maya: Two Routes to Local AI on Impossible Hardware

Salih M. October 27, 2026

https://gemini.google.com

The local AI world has a new family tree, and it’s worth understanding. Strata and Project Maya are two open-source inference engines designed to run models that normally have no business running on consumer hardware. They share DNA, but they aim at different users and different machines.

Strata is the more polished, general-purpose local runtime. It is best known for running Qwen3.8-Flash-Next, a 125B-parameter mixture-of-experts model with only 6B active parameters, on everything from an M5 Max MacBook to an RTX 3090. It uses custom Metal and CUDA kernels, streams cold experts and the model’s large n-gram table from SSD, and exposes an OpenAI/Anthropic-compatible API. Strata also has a desktop app, a terminal interface, and a growing set of prepacked quantizations. In practice, it makes a 125B model feel like a local chatbot rather than a science project. (Strata GitHub, Strata Mac docs)

Project Maya is the power-user cousin. It started from Strata’s engine and added its own quantization formats, model-specific tuning, and a separate DeepSeek engine called maya-ds4. Maya’s focus is even more extreme: GLM-5.3-Flash (321B) and DeepSeek-V4.1-Flash-1017 (748B) on one to sixteen GPUs, with NVIDIA and AMD support on Linux and Windows. It ships with an engine, server, dashboard, OpenAI/Anthropic APIs, and quant tools. It is not really a Mac app; it is a “put a huge model on a PC GPU farm” toolkit. (Project Maya GitHub, MindStudio explainer)

	Strata	Project Maya
Role	Polished local inference app	Power-user engine for massive MoE models
Main models	Qwen3.8-Flash-Next, DeepSeek V4.1-Flash	GLM-5.3-Flash 321B, DeepSeek V4.1-Flash-1017
Hardware	Apple Silicon, NVIDIA, some AMD	NVIDIA/AMD GPUs, one to 16 cards
Mac support	Yes, native Metal	Not a Mac-first project
Interface	Desktop app, terminal, API	Engine, server, dashboard, API
Memory trick	Hot experts in RAM; cold experts and n-gram table streamed from SSD	Same idea, tuned further for very large MoE models
License	MIT	MIT

The overlap is DeepSeek V4.1-Flash. Strata now supports it on PC, but not on Mac yet. Maya supports it across a wider range of GPU and multi-GPU setups, with its own Maya-XS and Maya-S quantizations. If you have a single RTX 3090 and 64GB of RAM, Maya can run DeepSeek at roughly 8–17 tok/s depending on cache warmth and context—not fast, but genuinely local. Strata is the easier choice for Qwen Flash Next on a Mac or a single-GPU PC.

Notable similar software

Strata and Maya are not the only games in town. If you’re building a local AI stack, these are the names worth knowing:

Tool	What it brings	Why it matters
llama.cpp	The foundational GGUF inference engine	Broad hardware and model support; the common ancestor of many specialized runtimes
Ollama	Easy local model runner	The “just works” option for most home users, with a huge model library
LM Studio	Desktop GUI for local models	Popular, friendly, and good for model discovery and chat
MLX / mlx-lm	Apple’s ML framework	The foundation for most fast Apple Silicon inference
KTransformers	CPU/GPU hybrid inference for huge MoE models	One of the closest serious alternatives for DeepSeek-class models on consumer hardware
vLLM / SGLang	High-throughput serving engines	The server-grade comparison point for multi-GPU and production inference
oMLX / MTPLX	Apple Silicon runtimes	Mac-specific alternatives worth watching for Qwen/Flash-Next workloads

If you only add five names to the conversation, make them llama.cpp, Ollama, LM Studio, MLX, and KTransformers. Those give readers the full map without turning the article into a catalog.

Bottom line

Use Strata if you want the easiest path to a fast, local Qwen model on a Mac or a single machine. Use Maya if you have a PC with serious GPU and RAM capacity and want to run DeepSeek or GLM at home. They are not rivals so much as two branches of the same idea: keep the model local, stream what doesn’t fit, and make impossible hardware do useful work.

Notable similar software

Bottom line

References

You should also read: