Project Maya GitHub: Datacenter Models on Your Own GPUs

What is Project Maya, and where is it on GitHub?

Project Maya is a free, MIT-licensed engine that runs GLM-5.3-Flash (321 B), DeepSeek-V4.1-Flash-1017 (748 B), MiniMax-M3.1-Flash (428 B) and CYBER-FROST, a fine-tune of Qwen3.8-Flash-Next, on your own GPUs, from one card to 16. The Project Maya GitHub repository is mw00/project-maya, from Peasant Smith (@mw00), co-maintained with @needmorevram. It keeps the experts your conversation uses most on the GPU, the next ones in RAM and the rest on an NVMe SSD, and moves them as you chat.

The engine is a small download. The models are not: every build below is a separate download of 93.00 to 508.38 GB, and which one your PC can hold is the decision to make before you start it. This page covers Maya 1.1.51, tagged on 3 November 2026. Our Maya calculator is read at 1.1.51 too: it carries the GLM Maya-XS row (73.4 GB) and the DeepSeek Maya-XS v2 row (278.12 GB with the shared Q6_K tables). The 1.1.45 to 1.1.51 changes are in the section near the end. Maya ships several times a day, so every figure here carries its version.

Source: Project Maya README, "What you need" and "Install".

Every Project Maya GitHub build, sized for one PC

Our calculator carries 27 Project Maya GitHub builds across four model families and sizes every one. Each table gives the download, the VRAM plus RAM Maya plans for where its files state it, and the verdict our Maya calculator gives one common machine: an RTX 4090 24 GB with 128 GB of RAM, at 128K of context on Linux. "Runs" means every expert fits in VRAM and RAM. "From the SSD" gives how much of the model is read from the disk while it answers: it still runs, slower.

DeepSeek V4.1 Flash 1017 (748 B)

Sources: Maya's tools/ds4/download_gguf.json at v1.1.51; downloads summed from the Hugging Face file headers in our calculator's data; verdicts from the live calculator, 3 November 2026.

DeepSeek's downloads include its Engram tables, 110.60 to 161.28 GB per build, which are read a few rows a token from the SSD and never count against VRAM or RAM. That is why a 276.63 GB Maya-XS fits a 24 GB card with 128 GB of RAM. The two abliterated mixes are orcarouter's community builds with refusals removed; the calculator sizes their memory but does not yet price their VRAM and RAM split.

Maya-XS v2, 117 GB, is the model files, not the whole build. The 1.1.47 release replaces Maya-XS with a version whose model files are 117 GB, while the shared 161 GB of Q6_K n-gram tables stay on disk. The total download therefore stays about 276.63 GB, and the calculator's Maya-XS row does not move to 117 GB. The update asks whether to replace an existing Maya-XS; it downloads only the 117 GB of new model files, keeps the tables, and removes the old files once the new ones are complete.

GLM-5.3-Flash (321 B)

Source: our calculator's data at Maya 1.1.51; verdicts from the live calculator, 3 November 2026.

GLM-5.3-Flash is the model Maya was built around: 321.3 billion parameters, 18 billion active per token, 288 experts with 8 in use. On this machine six of its eight builds fit in memory, Maya-L included. Our own run of an earlier release is on the GLM 5.3 Flash page.

MiniMax-M3.1-Flash (428 B)

Source: our calculator's data at Maya 1.1.51; verdicts from the live calculator, 3 November 2026.

CYBER-FROST, a Qwen3.8-Flash-Next fine-tune

Source: our calculator's data at Maya 1.1.51; verdict from the live calculator, 3 November 2026.

Maya plans the Qwen family's context by the VRAM: the KV cache may take about 12% of it, or about 6% when VRAM and RAM together fall short of the 84 GB the model needs.

Check your machine before you clone Project Maya from GitHub

Maya's setup works out the fit once it is installed. Our Maya calculator does it before you clone anything: it ports Maya's own placement rules at 1.1.51, puts each build's experts in your VRAM, RAM and SSD, and gives the verdicts above for your own cards, RAM, context and system. A Project Maya GitHub download of 339.48 GB is worth a minute of sizing first.

Two DeepSeek builds stand out for a PC with limited room:

• Short of disk: Unsloth's UD-IQ1_S is the smallest standard build, 232.98 GB, which is 43.65 GB under Maya-XS. Its Engram tables are stored as IQ4_NL instead of Maya's Q6_K. It is set up by flag, --ds4-quant gguf-iq1_s.
• Short of RAM: Maya-XS needs 121 GB of VRAM plus RAM against 123 GB for UD-IQ1_S, because its network is 7.04 GB smaller. Maya's README describes it as the whole model in a 128 GB PC's memory, with the tables on the SSD.

Which DeepSeek build stays closest to the original?

Maya's 1.1.43 notes measure it. KLD against the original model, where lower means the next-token choices drift less:

Source: Project Maya v1.1.43 release notes.

Maya-XS v2, in 1.1.47, measures 0.251 against 0.284 for the version it replaces, so it closes about 12% of the gap to the original. It is still above Maya-S's 0.119, and the 117 GB figure is the model files, not the whole download.

UD-IQ3_S sits closer to Q8 than to Maya-S, and the cost is size: 423.63 GB of files against 339.48 for Maya-S. Maya's README gives a second measure for Maya-S, the same next token as the original 92.6% of the time. Both measures come from the maintainers, on their own test sets. The notes give no KLD for the 1-bit builds; Strata's 0.2.9 release measured UD-IQ1_S and UD-IQ1_M, and our Strata article sets out that table.

How fast do people measure it?

Maya's README collects reports from its users, most from the built-in --bench command, with the version named. These DeepSeek rows are theirs, not ours.

Source: Project Maya README, "How fast is it?"; the last row is @axitek's chart, read on X, 2 November 2026.

The RTX 4090 row is the one most readers will resemble: one consumer card with plenty of RAM, on Maya-XS, writing at 44.1 tokens/s and reading a prompt at 539. Version 1.1.42 adds that a fresh 20-30K-token DeepSeek prompt reads 16-24% faster than in 1.1.41.

The last row is the decode story with a warning. @axitek's 2x RTX 3090 machine moved from 25 to 30 tokens/s on the old Maya-XS to 40 to 50 on Maya-XS v2, and the chart reads 43.4 at 262K. Prefill moved the other way: about 1,220 to 1,500 on 1.1.28 became about 950 to 1,070 on 1.1.48. If your work is long prompts rather than long answers, the v2 update can be slower on this class of machine.

Project Maya or Strata?

Both run DeepSeek V4.1 Flash 1017, and they are related: Strata's DeepSeek engine is ported from Maya's at release 1.1.17, and Strata's notice says "Project Maya is itself built on Strata." Pick by your machine and your model:

• GLM-5.3-Flash or MiniMax-M3.1-Flash: Project Maya. Strata does not run them.
• More than one card: Project Maya plans for one GPU up to 16.
• A Mac: Strata, for DeepSeek. Maya's Mac support covers GLM-5.3-Flash only.
• Intel Arc for DeepSeek: neither engine.

For DeepSeek on an NVIDIA or AMD PC, both work, with different menus. Size the same machine in the Maya calculator and the Strata calculator, and read our Strata article for its sets.

What changed between 1.1.42 and 1.1.44?

• Unsloth's DeepSeek files. Version 1.1.43 (31 October, 09:21 UTC) runs Unsloth's DeepSeek GGUF quants and put UD-IQ3_S in the setup list. Version 1.1.44 (12:50 UTC) listed UD-IQ1_M too. The others are set up by flag. Pictures work with every Unsloth quant, through a 0.97 GB encoder file from Unsloth's repository.
• One copy of the tables. Maya-S, Maya-XS and UD-IQ3_S hold the same 161 GB of Engram tables, so setting up a second one reuses the first (1.1.43).
• Long agent conversations. When the loop guard steps in, the prompt is kept whole. On a 150K-token conversation the next turn took about 15 times less time with GLM and with DeepSeek (1.1.43).
• Pictures on older cards. DeepSeek's EXL3 quants now read pictures on V100s too (1.1.42).

What changed between 1.1.45 and 1.1.51?

• 1.1.45, 31 October: a one-GPU DeepSeek bug could fail after several long prompts with an invalid memory access, because a prompt used expert memory the engine was still loading in the background. The load now finishes first. Two or more GPUs were not affected.
• 1.1.46, 1 November: DeepSeek reads a 16K prompt about 5% faster on Strix Halo and Gorgon Halo, with the same answers. Long agent sessions whose messages carry content as parts, and tool calls with an unclosed parameter tag, no longer fail the request.
• 1.1.47, 1 November: Maya-XS v2. KLD 0.284 to 0.251, about 20% faster writing, model files 117 GB. One set of Q6_K n-gram tables now serves every DeepSeek quant, so the Q8 download falls from 508 to 467 GB. The update keeps the tables and removes the old model files once the new ones are complete.
• 1.1.48, 1 November: DeepSeek on several GPUs reads prompts in 4K chunks by default. An 8K prompt is 24 to 30% faster on 3x V100, and about 3.5% more experts fit on the cards. A thinking-budget transition reuses cache instead of rereading the thinking, which was 59 seconds at 32K. A long prompt no longer stops an answer that is already writing.
• 1.1.49, 2 November: fixes the Maya-XS v1 to v2 update stopping with "changed after the pack was built", and gives back about 161 GB of old n-gram files. Qwen CYBER-FROST runs on 8 GB cards again.
• 1.1.50, 2 November: a Models page that lists every Maya model with its size and can download, load, unload and delete, plus a model API (/api/tags, /api/ps, /api/show, /api/chat, /api/generate, /api/pull, /api/delete). A request for a model that is not loaded swaps to it, and keep_alive unloads it when idle. Turbo is one button beside the model's parameters: greedy sampling, no thinking, a 32K context on GLM and MiniMax, expert skipping and the fastest drafting. It says "Turbo ON" while it is on, and switching it off restores your settings exactly; the long-form reports put the gain at up to about 24% faster decode. It trades quality for speed, except that on the smallest quants, GLM Maya-XS and DeepSeek Maya-XS, it keeps every expert. A thinking budget field sets how many tokens the model may think, default 32768 (8192 for GLM Maya-XS). GLM-5.3-Flash Maya-XS is a new, smaller GLM quant, about 73 GB, for a PC with 64 GB of RAM and a 24 GB card at 128K; setup offers it as glm-maya-xs with an int8 KV cache. DeepSeek long prompts read up to about 19% faster where RAM holds most of the experts. Maya-S reuses Maya-XS's n-gram tables instead of downloading a second pair.
• 1.1.51, 3 November: DeepSeek in tensor parallel fills every GPU with experts, following the room each card has left after the drafter, head and buffers, so no card keeps VRAM unused. On the machines it was measured on, about 10% more experts sit in VRAM and decode is up to 10% faster; DS4_TP_SHARES=off keeps the even split. With most experts in RAM on several GPUs, the prompt path borrows the experts' room and reads in larger pieces, so prompts from about 10K tokens read up to about 12% faster and about 7% more experts stay in VRAM, while prompts around 8K are about 3% slower on those setups; machines whose experts are mostly in VRAM are unchanged. CYBER-FROST gets a thinking Off button. Several conversations at once no longer disturb each other when one is stopped.

Two opt-in items are worth knowing before you turn anything on. Turbo is a quality trade, so benchmark with it off as well. DS4_AR_P2P=1 does peer-to-peer all-reduce over ReBAR for DeepSeek on two or more GPUs, reported up to 29% faster prompts on the PCs it was measured on, but it was a little slower on V100s, which is why Maya does not enable it.

The calculator is now read at 1.1.51. These release notes are read at source; the size tables and verdicts on this page are the calculator's, with the two new rows above.

FAQ

What is Project Maya?

A free, MIT-licensed local inference engine for GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, MiniMax-M3.1-Flash and CYBER-FROST, a fine-tune of Qwen3.8-Flash-Next. It spreads experts across GPUs, RAM and an NVMe SSD, and works on NVIDIA from the V100 up, with experimental AMD and Windows support.

Where is the Project Maya GitHub repository?

At github.com/mw00/project-maya. Releases come several times a day; this page covers 1.1.51, and our calculator's data is read at 1.1.51 too.

Which models does Project Maya run?

Four families: DeepSeek V4.1 Flash 1017 in 15 builds, GLM-5.3-Flash in 8, MiniMax-M3.1-Flash in 3, and CYBER-FROST, a Qwen3.8-Flash-Next fine-tune, in 1. Our calculator sizes all 27.

How much disk and memory do I need before downloading Project Maya's DeepSeek?

The builds weigh 224.57 to 508.38 GB, so plan that much free SSD space. Maya plans 121 GB of VRAM plus RAM for Maya-XS, 186 GB for the default Maya-S and 263 GB for UD-IQ3_S; with less, it reads the rest from the SSD and runs slower. Our Maya calculator checks your exact machine.

Which Project Maya GitHub build should a 24 GB card with 128 GB of RAM use?

On that machine at 128K, our calculator has Maya-XS, both Unsloth 1-bit builds and Mix-128G running DeepSeek in memory, every GLM-5.3-Flash build up to Maya-L and EXL3 3.05bpw, MiniMax's Maya-XS and Maya-S, and CYBER-FROST. The default DeepSeek Maya-S reads 25.7 GB from the SSD.

Does Project Maya run on Intel Arc or a Mac?

Intel Arc is not supported by Maya's engine. Apple Silicon is experimental and for GLM-5.3-Flash only, and Maya's authors have not tested it with the real model.

Is UD-IQ3_S better than Maya-S?

On Maya's KLD measure, yes: 0.061 against 0.119, close to Q8's 0.047. The cost is 423.63 GB of files against 339.48 GB, and 263 GB of VRAM plus RAM against 186 GB.