project-maya

GLM-5.3-Flash (321 B), DeepSeek-V4.1-Flash-1017 (748 B) and MiniMax-M3.1-Flash (428 B) on your own GPU
private and fast · one NVIDIA GPU or up to 16 · AMD and Windows (experimental) · chat, pictures, OpenAI- and Anthropic-compatible API · several requests at once

Created and led by Peasant Smith (@mw00), co-maintained with @needmorevram

_{Project Maya is free. If it runs well on your PC, sponsoring it pays for the test hardware and the work on it.}

Models this size usually need a datacenter. Maya runs three of them on the hardware you already have: GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017 and MiniMax-M3.1-Flash. Each model has an engine built for it. It keeps the experts your conversation uses most on the GPU, the next ones in RAM and the rest on your NVMe SSD, and moves them as you chat. It measures your GPUs, CPU, RAM and SSD and tunes itself to them. Nothing leaves your machine.

• Fast: up to 149 tokens/s on two RTX 5090s and an RTX 4090, 124 on two H100s, 118 on four RTX 4090s, 81 on one RTX PRO 6000, 30+ on a single RTX 5090. Real numbers from real users.
• Close to the original: Maya's own quants of GLM-5.3-Flash keep 97.7-99.2% of the FP8 model's zero-shot accuracy; DeepSeek V4.1 · Maya-S picks the same next token as the original 92.6% of the time, MiniMax M3.1 · Maya-S 87.6% (the average over its categories).
• Plugs into your tools: OpenAI- and Anthropic-compatible API with tool calls, ready for coding agents.
• Several requests at once: up to 8 answered together with GLM-5.3-Flash, DeepSeek-V4.1-Flash and MiniMax-M3.1-Flash - for agents that run several sessions, or several people on one PC.
• Everything in the box: a chat and a live monitor in the browser, pictures, thinking levels, a one-command setup.

git clone https://github.com/mw00/project-maya.git && cd project-maya
./setup.sh

The setup asks which model you want, downloads it, builds everything once and starts the model. Then open http://127.0.0.1:8080. Windows: START-MAYA.bat instead of ./setup.sh. More in Install.

Every Maya quant is named after its model, then its tier - XS, S, M or L, by how many bits its experts keep. The setup recommends one for your PC; to pick one yourself, give its setup id: ./setup.sh --model glm-maya-s24.

All of them run on NVIDIA from the V100 / RTX 20 on, and on AMD and Windows (experimental; see What you need). DeepSeek's sizes include its Engram tables. The setup also offers unsloth's GGUF and LibertAIDAI's EXL3 quants of DeepSeek, orcarouter's abliterated DeepSeek, Mia-AiLab's EXL3 quant of MiniMax, turboderp's EXL3 quants of GLM, and Qwen-3.8-Flash-Next (CYBER-FROST) as its Q5_K_M GGUF, with pictures (its settings). Each model's files, drafters, pictures and options: docs/MODELS.md.

Maya's quants are made from each model's released weights and measured against the original:

• GLM-5.3 · Maya quants keep 97.7-99.2% of the FP8 model's zero-shot accuracy and pick its next token 83-90% of the time (Maya-L: 99.2%, 90%). How they're made and measured: MAYA-S, MAYA-S24, MAYA-M, MAYA-L.
• DeepSeek V4.1 · Maya-S picks the same next token as the original 92.6% of the time.
• MiniMax M3.1 · Maya-S picks it 87.6% of the time (the average over its categories).

More on each model's card on Hugging Face.

Real machines measured by our users, most with ./maya.sh --bench (decode: three greedy 256-token answers; prefill: an 8K-token prompt unless the row says otherwise). Share yours: --bench, then --report, in a GitHub issue.

A token is about ¾ of a word. The speed holds as a conversation grows, and the first answers after a start are the slowest while the caches fill with what you use.

• In the browser: chat (kept in your browser, with search, pins and folders, per-chat instructions and thinking level, edit and regenerate, HTML previews), a live Monitor of the model and your hardware (down to what fills each GPU's memory), a context size and port you can change in Settings, and Models: every model set up on this machine, switching to another (Maya starts again with it) and each one's engine settings (context, GPUs, drafting, images, memory), saved into its maya-<model>.json.
• Saved: the sidebar's Saved line adds up what the model's own API (DeepSeek's, Z.ai's or MiniMax's) would have charged for the same tokens - the prompt Maya reuses from its cache at the API's cached price, DeepSeek's peak hours at its peak price - kept per model in token-savings.json; its prices can be changed in Settings › Savings.
• Your apps and coding agents: an OpenAI-compatible API at http://127.0.0.1:8080/v1, and Anthropic's at http://127.0.0.1:8080/v1/messages (Claude Code: ANTHROPIC_BASE_URL=http://127.0.0.1:8080). Tool calls are streamed as the model writes them, and a long prompt's reading progress too when a request asks for it ("return_progress": true). A web page served from another address can call it too once you allow its address ("cors_origins", docs/SETUP.md).
• Responses API: /v1/responses supports text/image conversations, streaming, function tool calls, schema-validated output, background tasks with cancellation/replay, and optional shared in-memory history (store:true, previous_response_id). Responses are not stored by default. Options and limits: docs/RESPONSES.md.
• Thinking: Off, Low, High (the default) or Max, in Settings or per request.
• Pictures: attach one in the chat, or send image parts through either API.
• From another device: ./maya.sh --setup --host 0.0.0.0 --api-key <secret>. Always set a key.
• Tuning: ./maya.sh --calibrate tunes the engine to your hardware in 10-15 minutes. Every setting: docs/SETTINGS.md.
• Each model's own options - GLM-5.3-Flash's several requests at once (--parallel), long contexts and DFlash2 drafter; DeepSeek's kept conversations, expert loading and Engram read-ahead; MiniMax's thinking levels: docs/MODELS.md.

The setup checks all of this, prints the exact command for anything missing, and installs nothing system-wide.

Run the two commands in Quick start and answer a few questions (Enter takes the recommended choice). The setup shows every download before it starts it, builds everything once (20-40 minutes plus the download) and starts the model. Then open http://127.0.0.1:8080. Ctrl+C stops it, and ./setup.sh starts it again right away.

• Windows: START-MAYA.bat instead of ./setup.sh - see docs/WINDOWS.md.
• AMD: docs/AMD_MAYA.md.
• Mac with Apple Silicon (experimental, GLM-5.3-Flash): ./maya.sh --backend metal - see docs/METAL.md. Built and unit-tested on GitHub's Macs, not yet tested by us with the real model (we have no Mac): if you run it, please tell us how it went in an issue.
• Intel Arc: not supported by Maya's engine; a community report of the Maya-XS DeepSeek GGUFs on llama.cpp's SYCL backend: docs/INTEL_ARC.md.
• Updating: the page's Settings › About and updates, or git pull then ./setup.sh. The model isn't downloaded again.
• Model files downloaded another way (a download manager, another PC, a Hugging Face cache): ./maya.sh --import-model FOLDER finds Maya's files in FOLDER by size and checksum, whatever they are called, and moves them into place (--copy when FOLDER is on another drive; the originals stay). Files that do not match are left alone.
• Every option (GPUs, context, a model folder you already have, remote access, ...): docs/SETUP.md.

The start takes a few minutes: the engine pins most of the free RAM (all but about 6 GB) for its expert tier, so other programs get little RAM while Maya runs.

Run ./maya.sh --report (Windows: START-MAYA.bat --report) right after a slow answer or an error, and attach the maya-report.txt it writes to a GitHub issue. It holds your hardware, your setup and the engine's speed lines; nothing is sent anywhere by itself. The engine's own log is maya-<model>.log in the Maya folder.

• The build stops ("nvcc ... cannot build for these GPUs", "unsupported GNU version", or cuda_profiler_api.h / nvtx3/nvToolsExt.h missing): docs/SETUP.md has the fix for each.
• Slow, with the SSD busy all the time: not enough free RAM for the experts. Close programs or add RAM.
• Windows: answers suddenly several times slower: the card is too full and Windows moves GPU memory to RAM. Leave more of it free: START-MAYA.bat --setup --vram-reserve-mib 2048 (the expert pool's size).
• Windows on a Strix Halo / Gorgon Halo: other programs run out of memory, or the setup warns about the page file: set the page file to the size the setup prints (docs/AMD_MAYA.md).
• The download stopped: run ./maya.sh again; it continues where it stopped.
• Port 8080 is in use: Maya is already running, or --port 8081 starts another one. Settings > Port moves Maya to another port for good (it starts again there).

Project Maya is free and open source. Every release is tested on our own NVIDIA and AMD machines before it ships (the Mac support so far only on GitHub's virtual Macs), and sponsorship pays for that test hardware (the current goal: 50 monthly sponsors for a real Mac and new GPUs, so Maya runs well on more hardware), for building new quants and models, and for faster fixes. If Maya saves you a cloud bill, consider sponsoring a small part of it back - thank you!

Thank you to everyone who sponsors Project Maya. Sponsors at $15 a month and up are listed here - become the first.

Speed reports, bug reports and pull requests are welcome: CONTRIBUTING.md says what helps most.

• People: Project Maya is created and led by Peasant Smith (@mw00, X), who owns the project, its name and the Maya models. @needmorevram (X) is a co-maintainer and contributed major features: DFlash2 drafting, EXL3 support, KV streaming, conversation checkpoints, the layer-split search, the expert profile and the CPU expert kernels, and for maya-ds4 the 3.5-bit EXL3 option, tensor parallel over 4 / 8 GPUs and the threaded expert-cache swaps (#276). Thanks also to everyone who sent code and speed reports (contributors).
• Built on Strata (MIT License, Copyright (c) 2026 Niko1221 and the Strata contributors): Maya's engine started from Strata's and was rewritten for GLM-5.3-Flash (the expert tiers across VRAM, RAM and SSD, the layer split, MTP decoding), and its server and dashboard grew from Strata's. And on ggml / llama.cpp (MIT License, Copyright (c) 2023-2026 The ggml authors): the quantization formats, the CPU dot products and the prompt path's MMQ kernels, built from a pinned commit (third_party/ggml/LICENSE).
• EXL3 support follows exllamav3's format, codebooks, CPU expert kernel and tensor-core decoding of a tile (MIT License, Copyright (c) 2025 Turboderp; third_party/exllamav3/LICENSE). The DFlash2 drafter's forward follows beellama.cpp's (MIT License, Copyright (c) 2023-2026 The ggml authors, Anbeeld; third_party/beellama.cpp/LICENSE).
• The models: GLM-5.3-Flash by zai-org (MIT); Maya's quants and the image encoder file are made from zai-org's released weights and keep its license. DeepSeek-V4.1-Flash-1017 by DeepSeek (MIT); Maya's quant, its Engram tables and the drafter are made from DeepSeek's released weights and keep its license; the vision tower comes from DeepSeek's checkpoint, the GGUF option from unsloth's GGUFs, the EXL3 options are LibertAIDAI's (2.9 and 3.5 bpw). The optional DFlash2 drafter (GLM-5.3-Flash-DFlash2-GGUF by Anbeeld) is CC BY-NC-ND 4.0, non-commercial use only, and is downloaded only when you choose it.
• The dashboard's font: Outfit (SIL Open Font License 1.1, serve/web/fonts/OFL.txt); its icons: Lucide (ISC License, serve/web/LICENSE.lucide.txt); its UI library: Preact (MIT License) with htm (Apache License 2.0), bundled in serve/web/preact.js with their notices.
• Maya is open source under the MIT License; the notices of Strata, ggml, exllamav3 and beellama.cpp stay with every copy.