project-maya

Run GLM-5.3 (744B), GLM-5.3-Flash (321B), DeepSeek-V4.1-Flash (748B) and MiniMax-M3.1-Flash (428B) on your own GPUs - one engine, one dashboard, OpenAI- and Anthropic-compatible API.

Automatically checked · GitHub checked 2026-11-12

project-maya

Overview

English | 日本語 | 简体中文 | Português

Maya logo

Project Maya

GLM-5.3 (744 B), 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 · chat, pictures, OpenAI- and Anthropic-compatible API · several requests at once

GLM-5.3 Maya downloads GLM-5.3-Flash Maya downloads GLM-5.3-Flash Maya-M-Derisked downloads DeepSeek-V4.1-Flash Maya downloads MiniMax-M3.1-Flash Maya downloads

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

Sponsor on GitHub Buy Me A Coffee
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 four of them on the hardware you already have: GLM-5.3, 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 on every model, thinking levels, a Models page that downloads, loads and deletes models, a Turbo button, a one-command setup.

Quick start

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.

All documentation: docs/README.md

Pick your model

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.

ModelMaya quant · setup idSizeMade for
GLM-5.3-Flash
zai-org · MIT · 321 B, 18 B active · up to 1M context
Maya-S (default) · glm-maya-s96.5 GBPCs with a smaller memory pool (RAM and VRAM together)
Maya-S24 · glm-maya-s2494.7 GBcards of 24 GB or less: ~1.5 GB more of the card for experts, decode +14%
Maya-XS · glm-maya-xs73.4 GBPCs with 64 GB of RAM and a 24 GB card, at 128K context
Maya-M · glm-maya-m116 GBa bigger memory pool; calibrated toward tool calls and front-end code
Maya-M-Derisked · glm-maya-m-derisked116 GBMaya-M with a weight edit by Blackfrost_AI that reduces blanket refusals; experimental
Maya-L · glm-maya-l156.3 GBthe biggest memory pools; the closest to the original
GLM-5.3
zai-org · its own license · 744 B, 40 B active · up to 1M context · text only
Maya-XS · glm-5.3-maya-xs170.0 GBthe whole model in a 192 GB memory pool (VRAM and RAM together); NVIDIA from the V100 on and AMD (details)
UD-IQ2_M (unsloth) · GLM-5.3-UD-IQ2_M238.6 GBthe biggest memory pools; closer to the original
DeepSeek-V4.1-Flash-1017
DeepSeek · MIT · 748 B
Maya-S (default) · deepseek-maya-s340 GBany amount of memory: what VRAM and RAM don't hold is read from the SSD
Maya-XS (v2) · deepseek-maya-xs277 GBthe whole model in a 128 GB PC's memory (the Engram tables stay on the SSD); v2 is about 12% closer to the original and about 20% faster at writing than v1 - running the setup again offers the update
MiniMax-M3.1-Flash
MiniMax · MiniMax Community License · 428 B, 23 B active · up to 1M context
Maya-XS · minimax-maya-xs93 GBthe whole model in a 128 GB PC's memory
Maya-S · minimax-maya-s122 GBcloser to the original

Which model fits your PC? See docs/MEMORY.md for each model's recommended memory (VRAM + RAM together).

All of them run on NVIDIA from the V100 / RTX 20 on, and on AMD and Windows (see What you need). DeepSeek's sizes include its Engram tables. The setup also offers unsloth's GGUF and turboderp's and LibertAIDAI's EXL3 quants of DeepSeek, orcarouter's abliterated DeepSeek, Mia-AiLab's EXL3 quant of MiniMax, turboderp's EXL3 quants of GLM, Qwen-3.8-Flash-Next (CYBER-FROST) as its Q5_K_M GGUF and Qwen-3.8-Flash-Next itself as unsloth's UD-IQ2_XS GGUF, both with pictures (its settings). Each model's files, drafters, pictures and options: docs/MODELS.md.

Quality

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

  • GLM-5.3-Flash · 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.

How fast is it?

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.

GLM-5.3-Flash speed benchmarks
MachineModelDecode (writing the answer)Prefill (reading your prompt)By
2x RTX 5090 32 GB + RTX 4090 48 GB (PCIe 5.0 x8, 5.0 x8, 4.0 x4), Core i7-13700K, 188 GB RAMMaya-S / Maya-M-Derisked (DFlash2), 32K context148.6 / 123.4 tokens/s2605 / 2341 tokens/s@prismaillya17 (v1.1.26, #498)
2x H100 80 GB (PCIe 5.0), Xeon Platinum 8488C, 504 GB RAMMaya-L / Maya-M-Derisked, 1M context122.4 / 123.7 tokens/s2099 / 1986 tokens/s@Molyleaf (v1.1.11, #258, #259)
4x RTX 4090 24 GB (PCIe 4.0 x16), Xeon w7-3445, 62 GB RAMMaya-S24118 tokens/s4129 tokens/s@0xPreDa (v1.0.31, #89)
8x RTX 3090 24 GB (PCIe 4.0, six x16 and two x8), EPYC 7402P, 126 GB RAMEXL3-4.05bpw (DFlash2), 1M context117.4 tokens/s1470 tokens/s@MarkMakers (v1.1.15, #414)
4x RTX 3090 24 GBMaya-S24 / Maya-M (DFlash2), 256K context100.8 / 86.5 tokens/s979 / 907 tokens/s@Whamp (v1.1.19, #349)
2x A100 80 GB (PCIe 4.0 x16), EPYC 7742 (a VM, 25 threads), 161 GB RAMMaya-M, 128K context91.9 tokens/s-@driona-will-fluff (v1.1.19, #326)
RTX PRO 6000 Blackwell 96 GB, Ryzen 7 9800X3D, 61 GB RAMMaya-M, 256K context81.0 tokens/s3102 tokens/s@yundddd (v1.1.7, #172)
4x Tesla V100 32 GB (PCIe 3.0 x16), 2x Xeon E5-2630 v4, 126-504 GB RAM, SATA SSDMaya-S / Maya-M / Maya-L, 32K context69.7 / 67.2 / 43.3 tokens/s730 / 713 / 627 tokens/s@HansStrupat (v1.1.36, #327, #340, #653)
4x Tesla V100-SXM2 32 GB (NVLink 2.0)Maya-S, 32K context68.4 tokens/s549 tokens/s@jeremiahlangner (v1.1.18, #333)
4x Tesla V100-SXM2 16 GB, Xeon E5-2643 v4, 63 GB RAMMaya-XS, 128K context64.0 tokens/s (after --calibrate)596 tokens/s@atollerus-tm (v1.1.63, #997)
1x H100 80 GB (PCIe 5.0 x16), Xeon Platinum 8488C, 504 GB RAMMaya-M-Derisked, 1M context61.4 tokens/s2014 tokens/s@Molyleaf (v1.1.15, #260)
2x NVIDIA CMP 170HX 64 GB (PCIe Gen2 x4 each), 2x Xeon E5-2690 v4, 91 GB RAMMaya-M, every expert in VRAM61.3 tokens/s514 tokens/s@ZackO2o (v1.0.6, #22)
2x RTX A6000 48 GB (PCIe 4.0 x8), Core i9-13900K, 128 GB RAM, SATA SSDMaya-S / Maya-M / Maya-L, 131K-token prompts60.7 / 50.9 / 32.1 tokens/s1726 / 1629 / 1298 tokens/s (131K)@everson (v1.1.20, #371)
4x RTX 4090 24 GB (PCIe 4.0 x16), Xeon Platinum 8380, 251 GB RAM, SATA SSDMaya-L, 1M context60.6 tokens/s1458 tokens/s (654 at 2K)@oscarlius (v1.1.57, #938)
2x RX 7900 XTX 24 GB (PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB)Maya-M, up to 512K context56.0 tokens/s (50.1 after a 456K-token prompt)907 tokens/s (114K)@pepuscz (v1.1.25 + #454-#457, #458)
2x RTX 5880 Ada 45 GB (PCIe 4.0 x16), Xeon Gold 5420+, 393 GB RAMMaya-L, 256K context54.5 tokens/s1401 tokens/s@vjsyong (v1.1.4, #149)
2x NVIDIA CMP 170HX 40 GB (PCIe Gen2 x16 each), Xeon w3-2423, 128 GB DDR5Maya-S, 128K context54.3 tokens/s831 tokens/s@nvsofts (v1.1.15, #263)
2x RTX 3090 24 GB (PCIe 4.0 x16, NVLink, power-capped at 252 W), Threadripper PRO 3975WX, 512 GB RAMMaya-M / Maya-L, 128K context54.1 / 40.3 tokens/s936 / 850 tokens/s@ksim63-gif (v1.1.23, #404)
2x RTX 3090 24 GB (PCIe 4.0 x16, power-capped at 225 W), Threadripper 3970X, 128 GB RAMMaya-S2453.0 tokens/s885 tokens/s@briepala (v1.1.2, #135)
3x Tesla V100 32 GB (PCIe 3.0 x16), 2x Xeon E5-2696 v3, 128 GB DDR3L, Windows 10Maya-S, 128K context52.9 tokens/s616 tokens/s@noahark (v1.1.20, #346)
RTX PRO 6000 Blackwell 96 GB (PCIe 5.0 x16), Ryzen 9 9950X3D, 253 GB RAM, Windows 11Maya-L, 512K context48.8 tokens/s3810 tokens/s@Andrey0710 (v1.1.43, #730)
4x Tesla V100 32 GB (PCIe 3.0 x16), 24 CPU cores, a 100 GiB RAM limitMaya-L, 32K context48.3 tokens/s (after --calibrate)499 tokens/s@areslp (v1.1.3, #130)
2x Radeon AI PRO R9700 32 GB (gfx1201), Core i5-13500, 96 GB DDR5Maya-S, 128K context45.2 tokens/s848 tokens/s@rim-jaynor (v1.1.57, #942)
2x Radeon AI PRO R9700 32 GB (gfx1201, PCIe 4.0 x16 each, 240 W), EPYC 7313, 256 GB RAMMaya-M-Derisked, 256K context44.4 tokens/s703 tokens/s@knoodrake (v1.1.28, #509)
4x RTX 3090 24 GB (PCIe 4.0 x8 behind one switch), Xeon Platinum 8260M, 376 GB RAMMaya-L, 1M context44.2 tokens/s (tuned)698 tokens/s@foureight84 (v1.1.17, #302)
2x Radeon AI PRO R9700 32 GB (gfx1201, PCIe 4.0 x8 each), Ryzen 5 3600, 64 GB RAMMaya-S24 / Maya-M-Derisked, 256K context44.1 / 24.9 tokens/s704 / 617 tokens/s@gegnep (v1.1.20, #395)
2x RTX 3090 24 GB (PCIe 4.0 x16, power-capped at 250 W), EPYC 7452, 256 GB RAMMaya-M / Maya-L, 128K context44.0 / 30.2 tokens/s860 / 801 tokens/s@orochifj (v1.1.4, #165)
8x AMD Instinct MI50 32 GB (gfx906, PCIe), EPYC 7K62, 112 GB RAMMaya-L, 128K context43.3 tokens/s288 tokens/s@abluesheep (v1.1.52, #830)
RTX PRO 6000 Blackwell 96 GB (PCIe 5.0 x16), Ryzen 9 9950X3D, 253 GB RAM, Windows 11Maya-L, a coding-agent session grown from 265K to 482K tokens42.4 tokens/s (median of 51 answers)-@Andrey0710 (v1.1.53, #873)
2x RTX 4090 24 GB (PCIe 3.0 + 4.0 x16), 2x Xeon Silver 4309Y, 188 GB RAMMaya-S2440.5 tokens/s1271 tokens/s@fuad00 (v1.1.0, #122)
RTX PRO 6000 Blackwell 96 GB (PCIe 4.0 x16), Ryzen 5 5600, 92 GB DDR4Maya-L, 1M context39.6 tokens/s2708 tokens/s (3246 warm)@QueTensors (v1.1.28, #532)
4x Tesla T10 15 GB (PCIe 3.0 x16), Xeon E5-2680 v4, 63 GB RAMMaya-S24, 100K context38.4 tokens/s393 tokens/s@zeallag (v1.1.16, #291)
2x Tesla V100 32 GB (PCIe 3.0 x16 each, power-capped at 150 W), EPYC 7V12, 256 GB DDR4-3200Maya-L, 128K context34.9 tokens/s (after --calibrate)589 tokens/s@scranst (v1.1.52, #939)
2x Tesla V100 32 GB (PCIe 3.0 x8 each), Ryzen 9 7945HX3D, 96 GB RAMMaya-M / Maya-L33.8 / 13.3 tokens/s490 / 372 tokens/s@nkhaooat (v1.1.11, #208)
RTX 3090 24 GB + RTX 3060 12 GB (PCIe 4.0 x16), EPYC 7702P (8-channel DDR4), 141 GB RAM (a VM)Maya-M, 128K context33.7 tokens/s682 tokens/s@erikash (v1.1.4, #157)
2x RX 9070 16 GB (gfx1201, PCIe x8 each), Ryzen 9 9950X, 89 GB RAMMaya-S24, 128K context, tuned33.0 tokens/s562 tokens/s@bob-u (v1.1.56, #934)
AMD Radeon 8065S (Ryzen AI Max+ PRO 495, 192 GB, 160 GB of it for the GPU), Windows 11Maya-S, 32K context31.5 tokens/s342 tokens/s@tobitege (v1.1.37, #663)
RTX 5090 32 GB, Ryzen 9 7950X, 96 GB DDR5Maya-M30+ tokens/s~2000 tokens/s@siganos (#43)
2x RTX 3090 24 GB (PCIe 4.0 x16, NVLink), Threadripper PRO 3955WX, 247 GB RAMEXL3-3.05bpw, 128K context29.2 tokens/s-@catonooka (v1.1.9, #187)
2x Tesla P40 24 GB (Pascal, PCIe 3.0 x16), EPYC 7302P (a VM), 184 GB RAMMaya-S24 / Maya-L28.0 / 16.9 tokens/s202 / 194 tokens/s@Jw88888 (v1.1.11 + #214, #212)
RX 7900 XTX 24 GB + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB)Maya-S24, 114K-token prompts27.5 tokens/s836 tokens/s (114K)@pepuscz (v1.1.17, #311)
1x RTX 3090 24 GB (PCIe 4.0 x16), EPYC 7702P (8-channel DDR4), 141 GB RAM (a VM)Maya-M, 128K context27.0 tokens/s1069 tokens/s@erikash (v1.1.4, #157)
RTX 4080 SUPER 32 GB (a 32 GB modded card), Threadripper 3970X, 157 GB RAM, SATA SSDMaya-S24, 262K context26.7 tokens/s1552 tokens/s@Micro-ATP (v1.1.11, #200)
3x Tesla V100 16 GB (PCIe 3.0 x16), 2x Xeon E5-2690 v4, 62 GB RAMMaya-S, 128K context26.6 tokens/s497 tokens/s@saatvik-viniak (v1.1.11, #211)
4x RTX 3080 20 GB, 64 GB DDR4Maya-M, 512K context26.3 tokens/s (a day of use)-@CoolJohn-lab (#104)
1x RTX 3090 24 GB (power-capped at 225 W), Threadripper 3970X, 128 GB RAMMaya-S2426.0 tokens/s913 tokens/s@briepala (v1.1.2, #135)
1x L40S 48 GB (vGPU), Xeon Platinum 8580 (a VM, 32 vCPUs), 504 GB RAMUD-Q4_K_XL / UD-Q5_K_XL / Q8_0, 512K context26.0 / 22.3 / 17.8 tokens/s1987 / 1877 / 1750 tokens/s (185K)@talisp (v1.1.32, #577)
RTX 4090 24 GB (PCIe 4.0 x16), 2x Xeon Silver 4310, 251 GB RAMMaya-S24, 32K context25.7 tokens/s1963 tokens/s@oscarlius (v1.1.24, #420)
RTX PRO 4000 Blackwell 24 GB, Core i9-13900K, 128 GB DDR5Maya-S2424.7 tokens/s1146 tokens/s@morrisfamily (v1.1.7, #174)
2x Tesla V100 16 GB (PCIe 3.0 x16), 2x Xeon E5-2696 v4, 128 GB DDR3-1866Maya-M, 256K context24.6 tokens/s (normal sampling)1092 tokens/s (12K)@c200762620-max (v1.1.49, #790)
RTX 5090 32 GB (PCIe 5.0 x16), Core i9-13900K, 64 GB RAM, Windows 11Maya-S, 128K context23.0 tokens/s2173 tokens/s@gianbing (v1.1.3, #139)
4x Tesla T4 16 GB, Xeon E5-2660 v2, 220 GB RAMMaya-S22.6 tokens/s416 tokens/s@ksanislo (v1.0.16, #36)
2x RTX 2080 Ti 22 GB (modded, NVLink, PCIe 3.0 x16, 250 W), Xeon E5-2686 v4, 247 GB RAM, SATA SSDMaya-M, 1M context22.0 tokens/s533 tokens/s@blackxwhite88 (v1.1.36, #675)
Radeon AI PRO R9700 32 GB (gfx1201), Core i5-14400F, 128 GB RAMMaya-S21.4 tokens/s-@potatopote-debug (v1.1.12, #213)
1x NVIDIA CMP 170HX 64 GB (PCIe Gen2 x16), Core i7-9700K, 32 GB RAMMaya-S24, 32K context20.6 tokens/s797 tokens/s@Kanorin-chan (v1.1.8, #281)
2x Quadro RTX 5000 16 GB (Turing, PCIe 3.0 x16 each, 200 W), Xeon W-2133, 94 GB RAMMaya-S, 128K context18.8 tokens/s494 tokens/s@arpleasa (v1.1.39, #707)
AMD Radeon 8060S (Ryzen AI Max+ 395, 128 GB, 96 GB of it for the GPU), Windows 11Maya-S18.3 tokens/s320 tokens/s@LIN-LAB-AI (v1.0.30, #102)
RTX 3090 Ti 24 GB (PCIe 3.0 x16, a VM), EPYC 7302P, 240 GB RAMMaya-L, 32K context18.2 tokens/s (tuned, #177)791 tokens/s@Jw88888 (v1.1.7, #177)
3x Tesla P100 16 GB (Pascal; PCIe 3.0 x8, 3.0 x8, 2.0 x4), Ryzen 5 2600, 64 GB RAMMaya-S24, 64K context17.9 tokens/s105 tokens/s@maxlawrie (v1.1.28 + #479, #504)
2x RX 6800 16 GB (gfx1030, PCIe 3.0 x16 each), Core i9-7980XE, 256 GB DDR4-2666Maya-L, 1M context16.7 tokens/s195 tokens/s@digitalhomesteader (v1.1.56, #895)
RTX 4090 24 GB, Ryzen 7 9700X, 64 GB RAMMaya-S24, 32K context16.6 tokens/s1320 tokens/s@MagikEh (#274)
RTX 5070 12 GB (PCIe 4.0 x16), EPYC 7302 (4-channel DDR4), 126 GB RAMMaya-S24, 64K context16.6 tokens/s323 tokens/s@hugou74130 (v1.1.45, #755)
NVIDIA DGX Spark (GB10, 121 GB unified memory)Maya-S, 128K context16.5 tokens/s~730 tokens/s (6K prompt)@ericoh88 (v1.1.15, #282)
RTX 4090 24 GB, Ryzen 9 9950X3D, 192 GB DDR5, Windows 11Maya-L, 128K context16.4 tokens/s1371 tokens/s@npc97 (v1.0.24, #63)
RTX A4000 16 GB (power-capped at 110 W), Core i7-14700, 64 GB DDR5-5600Maya-XS, 256K context, a real coding session15.2 tokens/s (15.6 at 32K)573 tokens/s (21K-token prompt)@kai-r (v1.1.64, #999)
RTX 5090 Laptop 24 GB, Core Ultra 9 275HX, 64 GB RAM, Windows 11 (a laptop)Maya-S2414.5 tokens/s994 tokens/s@klvnblst (v1.0.30, #67)
2x RTX 4060 Ti 16 GB (PCIe 3.0 x8 each), Core i7-10700K, 128 GB DDR4, SATA SSDUD-IQ3_XXS, 128K context13.0 tokens/s385 tokens/s@oisee (v1.1.42, #723)
DeepSeek-V4.1-Flash speed benchmarks
MachineModelDecode (writing the answer)Prefill (reading your prompt)By
8x RTX 5090 32 GB, Xeon Gold 6530, 1 TB RAMMaya-S, 512K context, --parallel 4184.2 tokens/s2510 tokens/s@etevrfed34-commits (v1.1.34, #585)
10x RTX 3090 24 GB (tensor parallel 2, PCIe 4.0 x4 each, 200 W), Xeon (24 threads), 64 GB RAMEXL3 2.9 bpw, 1M context97.0 tokens/s2390 tokens/s (128K)@DoktorMincs (v1.1.28, #494)
RTX 6000D 85 GB (PCIe 5.0 x16), Threadripper PRO 9965WX, 251 GB RAMEXL3 2.9 bpw, 512K context89.3 tokens/s1687 tokens/s (2836 at 33K)@Ross-Yang-94 (v1.1.25, #423)
2x A100-SXM4 80 GB, Xeon Platinum 8360Y, 1 TB RAMMaya-S, 1M context87.4 tokens/s2084 tokens/s@SunYong0821 (v1.1.34, #445, #594)
2x RX 7900 XTX 24 GB (tensor parallel 2, PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on a USB4 NVMeMaya-XS (v2), 512K context, DS4_INDEX_TC=182.5 tokens/s1587 tokens/s (130K; 663 at 8K)@pepuscz (v1.1.50 + #801, #797)
RTX PRO 6000 Blackwell 96 GB (PCIe 5.0 x16), Ryzen 9 9950X3D, 253 GB RAM, Windows 11Maya-S, 256K context72.3 tokens/s1757 tokens/s@Andrey0710 (v1.1.43, #737)
2x RX 7900 XTX 24 GB (tensor parallel 2, PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on a USB4 NVMeturboderp EXL3 2.25 bpw, 1M context, DSpark69.4 tokens/s1801 tokens/s (120K; 932 at 8K)@pepuscz (v1.1.56 + #867, #904, #945, #967, #968, #969)
2x RTX PRO 5000 Blackwell 72 GB, Ryzen 9 9950X3D, 249 GB DDR5Maya-S, 512K context68-71 tokens/s (prose), up to 162 (code), tuned~3700 tokens/s (42K)@realsquall1980-art (v1.1.28, #424)
4x RTX 4090 24 GB (PCIe 4.0 x16), 2x Xeon Platinum 8380, 251 GB RAM, SATA SSDMaya-S, 256K context67.5 tokens/s1087 tokens/s (400 at 2K)@oscarlius (v1.1.51)
1x A100-SXM4 80 GB, Xeon Platinum 8360Y, 1 TB RAMMaya-S, 1M context60.2 tokens/s241 tokens/s@SunYong0821 (v1.1.34, #594)
4x AMD Instinct MI100 32 GB (gfx908, PCIe 4.0 x16 each), EPYC 7402, 252 GB RAMMaya-XS, 32K context51.8 tokens/s172 tokens/s@nsaxploit (v1.1.22, #396)
4x RTX 3090 24 GB (PCIe 3.0, SM clocks capped at 1500 MHz), Threadripper 1950X, 126 GB RAMMaya-S, 256K context50-53 tokens/s (code), 40-43 (prose)947 tokens/s (10K), 1954 (90K)@mgabor3141 (v1.1.26, #448)
2x NVIDIA A40 48 GB (PCIe 4.0 x16), 2x Xeon Gold 6330, 256 GB RAM, Windows Server 2022Maya-S, 256K context48.9 tokens/s (tuned)~660 tokens/s (3K)@yjyz1011 (v1.1.28, #508)
2x RX 7900 XTX 24 GB (tensor parallel 2, PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on a USB4 NVMeturboderp EXL3 2.25 bpw, 1M context, DSpark48.6 tokens/s650 tokens/s (8K)@pepuscz (v1.1.56, #905)
2x RTX PRO 6000 Blackwell 96 GB (tensor parallel 2, PCIe 5.0), Core Ultra 7 270K Plus, 188 GB DDR5UD-Q8_K_XL, 1M context, DSpark48.3 tokens/s1245 tokens/s (3360 at 48K)v1.1.51, #819
RTX 4090 24 GB (PCIe 4.0 x8), Ryzen 9 9950X3D, 192 GB DDR5Maya-XS, 64K context44.1 tokens/s539 tokens/s@naeq (v1.1.27, #503)
4x Tesla V100 32 GB (PCIe 3.0 x16), 2x Xeon E5-2630 v4, 504 GB RAMMaya-S / UD-Q4_K_XL / UD-Q8_K_XL, 64K context44.0 / 28.6 / 29.7 tokens/s592 / 236 / 233 tokens/s@HansStrupat (v1.1.49, #785)
RX 7900 XTX 24 GB (PCIe 4.0 x4) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on a USB4 NVMeturboderp EXL3 2.25 bpw, 1M context, DSpark43.3 tokens/s371 tokens/s (8K)@pepuscz (v1.1.56, #908)
1x L40S 48 GB (vGPU), Xeon Platinum 8580 (a VM, 32 vCPUs), 504 GB RAMUD-Q8_K_XL, 256K context43.2 tokens/s1862 tokens/s (174K)@talisp (v1.1.32, #577)
AMD Radeon 8065S (Ryzen AI Max+ PRO 495, 192 GB, 160 GB of it for the GPU), Windows 11orcarouter abliterated Mix-128G, 256K context, DSpark43.2 tokens/s453 tokens/s@tobitege (v1.1.37, #663)
2x RX 7900 XTX 24 GB (tensor parallel 2, PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on a USB4 NVMeMaya-S, 512K context, DS4_INDEX_TC=141.7 tokens/s (49.4 after a 130K-token prompt, cached)1378 tokens/s (130K; 478 at 8K)@pepuscz (v1.1.48 + #801, #797)
4x RTX 3090 24 GB (PCIe 4.0 x16), Threadripper PRO 3975WX, 502 GB RAMUD-Q8_K_XL, 256K context37.1 tokens/s596 tokens/s@bandito1 (v1.1.50, #766)
4x RX 7900 XT 20 GB (gfx1100, --tp 4, tuned), 2x Xeon E5-2699C v4 (a Dell T7910), 1 TB DDR4-2400UD-Q4_K_XL35.0 tokens/s-@Cass67 (v1.1.32, #566)
2x Radeon AI PRO R9700 32 GB (gfx1201, PCIe 3.0 x16 each), Xeon E5-2690 v4, 126 GB RAMMaya-XS, 256K context34.0 tokens/s444 tokens/s@imaginitivename (v1.1.55, #890)
2x RTX 3090 24 GB (PCIe 4.0 x8, 270 W), Core Ultra 7 265K, 192 GB DDR5Maya-S, 1M context, tuned33.3 tokens/s467 tokens/s@Jellepepe (v1.1.35, #614)
2x RX 7900 XTX 24 GB (tensor parallel 2, PCIe 4.0 x4 each) + Radeon 8060S helper (Ryzen AI Max+ 395, 128 GB), the model on the internal NVMeLibertAIDAI EXL3 2.9 bpw, 1M context, DSpark32.6 tokens/s1455 tokens/s (160K; 376 at 8K)@pepuscz (v1.1.56 + #867, #904, #945, #967, #968, #987)
2x Tesla V100 16 GB (PCIe 3.0 x16), 2x Xeon E5-2696 v4, 128 GB DDR3-1866Maya-XS (v2), 256K context, DSpark32.0 tokens/s420 tokens/s@c200762620-max (v1.1.51, #815)
Tesla V100 32 GB (PCIe 3.0 x16, power-capped at 150 W), Xeon E5-2680 v4, 128 GB DDR4-2400Maya-XS (v2)24.5 tokens/s510 tokens/s after --calibrate (355 before)@wairentiun (v1.1.61, #978)
3x Tesla V100 32 GB (PCIe 3.0 x16), 2x Xeon E5-2696 v3, 128 GB DDR3L, Windows 10Maya-S, 64K context23-26 tokens/s (long answers)377 tokens/s@noahark (v1.1.26, #450)
2x RTX 2080 Ti 22 GB (modded, PCIe 3.0 x16, power-capped at 250 W), 2x Xeon E5-2696 v4, 94 GB RAMMaya-XS, 128K context23.1 tokens/s317 tokens/s@NorwinYu (v1.1.36, #659)
2x RTX 4090 24 GB (PCIe 3.0 + 4.0 x16), 2x Xeon Silver 4309Y, 188 GB RAMMaya-XS, 64K context21.5 tokens/s573 tokens/s@fuad00 (v1.1.22, #393)
RTX 4000 Ada 20 GB, Xeon w7-3455, 255 GB DDR5, Windows 11Maya-XS, 256K context21.1 tokens/s518 tokens/s (20K)@pietroMastro92 (v1.1.29, #534)
2x Tesla V100 32 GB (PCIe 3.0 x8 each), Ryzen 9 5900XT, 128 GB DDR4-3200Maya-S, 64K context20.8 tokens/s344 tokens/s@pirosap (v1.1.34, #586)
RX 7900 XTX 24 GB (gfx1100, PCIe 4.0 x16), Core i7-14700KF, 63 GB RAM, ROCm 10.2Maya-XS, 64K context, DSpark18.4 tokens/s395 tokens/s (146 at 2K)@Cayaf89 (v1.1.46, #757)
MiniMax-M3.1-Flash speed benchmarks
MachineModelDecode (writing the answer)Prefill (reading your prompt)By
RTX PRO 6000 Blackwell 96 GB (PCIe 5.0 x16), Ryzen 9 9950X3D, 253 GB RAM, Windows 11Maya-XS, 256K context, DSpark, refusal ablation on78.4 tokens/s1186 tokens/s (1332 at 2K)@Andrey0710 (v1.1.64, #1005)
RTX PRO 6000 Blackwell 96 GB (PCIe 5.0 x16), Ryzen 9 9950X3D, 253 GB RAM, Windows 11Maya-S, 256K context, DSpark48.6 tokens/s757 tokens/s@Andrey0710 (v1.1.43, #738)
RTX 3090 24 GB + 8x RTX 5060 Ti 16 GB (PCIe 3.0 x8 each), 183 GB RAMEXL3 3.5 bpw, tensor parallel 4 + DSpark, 8K context38.2 tokens/s (code), 32.4 (prose)-@needmorevram (#477)
1x L40S 48 GB (vGPU), Xeon Platinum 8580 (a VM, 32 vCPUs), 504 GB RAMMaya-S / EXL3 3.55 bpw, 256K context23.4 / 16.7 tokens/s332 / 273 tokens/s (158K)@talisp (v1.1.32, #577)
2x RTX 3090 24 GB (power-capped at 230 W), EPYC 7543 (a VM, 32 vCPUs), 224 GB RAMMaya-S, 200K context, DSpark22.0 tokens/s (code), 18.4 (prose)163 tokens/s (10K)@noonghunna (v1.1.30, #549)
AMD Radeon 8065S (Ryzen AI Max+ PRO 495, 192 GB, 160 GB of it for the GPU), Windows 11Maya-S, 128K context, DSpark21.0 tokens/s139 tokens/s@tobitege (v1.1.37, #663)
2x Tesla V100 16 GB (PCIe 3.0 x16), 2x Xeon E5-2696 v4, 128 GB DDR3-1866Maya-XS, 256K context, DSpark14.5 tokens/s254 tokens/s@c200762620-max (v1.1.51, #816)

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.

Using it

  • 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.
  • Models page: the catalog lists every Maya model with its size. Download one with a progress bar, load it, unload the running one or delete one - no command line needed.
  • Turbo: one button beside the model's parameters sets the fastest settings for the model (greedy sampling, no thinking, expert skipping, the fastest drafting). It trades some quality for speed; switching it off restores your settings exactly.
  • 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.
  • Model API: /api/tags, /api/ps, /api/show, /api/chat, /api/generate, /api/pull and /api/delete list, load, download and delete models from any app or script. A request for a model that is not loaded swaps to it, and keep_alive unloads it when idle. /models/load, /models/unload and /models/sse are on by default too ("model_management": false turns them off).
  • Thinking: Off, Low, High (the default) or Max, in Settings or per request, and a thinking budget field that sets how many tokens the model may think before it answers.
  • Loop guard: on by default - an agent that repeats the same tool call with the same result is steered off it.
  • Pictures: attach one in the chat, or send image parts through either API. Every model reads them (setup details per model: docs/MODELS.md). Videos too, for the models that read them.
  • Files: attach any file in the chat. Text and code go in as they are; PDFs as their text, or a scanned one's first pages as pictures; Word, Excel, PowerPoint, OpenDocument, EPUB and RTF files as their text. Anything Maya can't read says why. The same works through the APIs: OpenAI file / input_file parts and Anthropic document blocks (sent inline). poppler, when installed (sudo apt install poppler-utils), reads PDFs a little better; it is optional.
  • MCP tools: list MCP servers in the config ("mcp_servers"; Claude Desktop's mcpServers block works as it is) and the web chat's model can call their tools. A switch in the chat's + menu turns them on and off. What it does: serve/mcp.py.
  • Docker: there is no official image. docs/DOCKER.md is a recipe to build one from a release.
  • 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.

What you need

GPUNVIDIA, compute capability 7.0 or newer (V100, RTX 20 and newer): one, or up to 16 that share the model. More VRAM is faster. Pascal (Tesla P40 / P100, GTX 10) is experimental, not validated yet. AMD: RX 7900 XT / XTX, Radeon AI PRO R9700 / RX 9070, Strix Halo and Gorgon Halo (Radeon 8060S / 8065S); RX 6800 / 6900 and Radeon PRO V620 / W6800 not validated yet; GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 read pictures there too - docs/AMD_MAYA.md
RAM32 GB runs it. More RAM keeps more experts close and is faster
DiskA fast NVMe SSD with room for the model (~75 GB for GLM-5.3-Flash · Maya-XS, ~100 GB for GLM-5.3-Flash · Maya-S, ~160 GB for GLM-5.3-Flash · Maya-L, ~350 GB for DeepSeek V4.1 · Maya-S, ~470 GB for unsloth's Q8 GGUF of it, ~95 GB for MiniMax M3.1 · Maya-XS, ~200 GB for MiniMax-M3.1-Flash's EXL3 quant) and its picture encoder (1.1 GB). Not a hard disk
SystemLinux (x86-64; a CPU with AVX2 is best), the NVIDIA driver, CUDA 12.x (CUDA 13 works for RTX 20 and newer, not for the V100 or Pascal), g++, Python 3.10+. Windows 10/11: docs/WINDOWS.md. Not WSL2. AMD: ROCm 7

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

Install

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 only): ./maya.sh --backend metal (on a Mac, ./maya.sh picks Metal by itself) - 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.

Something went wrong?

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).

Your Support ❤️

Maya is free and open source, and it always will be. It exists so that anyone can run the biggest open models at home, on hardware they already own. Sponsors make that possible:

  • More hardware runs well. Every release is tested on our own NVIDIA and AMD machines before it ships. Macs are so far tested only on GitHub's virtual Macs. The current goal is 50 monthly sponsors, enough for a real Mac and newer GPUs.
  • New models sooner. Every new model and quant takes many hours of GPU time to build and check.
  • Faster fixes when something breaks on your setup.

If you enjoy running big models on your own machine, a small monthly sponsorship is the best way to keep Maya growing. Thank you!

Sponsor on GitHub Buy Me A Coffee

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

Credits and license

  • 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 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. GLM-5.3 by zai-org (its own license); its GGUF is unsloth's. 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 turboderp's (2.25 bpw) and 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 in LICENSE and third_party/ stay with every copy.