2026-10-20
Project Maya: GLM-5.3-Flash and DeepSeek with more memory and a low GPU bill
Maya quantizes GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 and serves them from one GPU, moving experts across VRAM, RAM and NVMe. The catalog profiles target one PRO 6000 and one H200.
Project Maya runs two open-weight MoE models on your own hardware: Z.ai's GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017. The code is MIT. Maya's quants keep the original weights' license, also MIT. SophyAI adds them as reference profiles next to the paths that already launch.
The quants
Maya publishes its own quants, keeping the MTP block that drafts tokens ahead.
| Quant | Model | Download | Fidelity Maya reports |
|---|---|---|---|
| Maya-M | GLM-5.3-Flash | 116 GB | 97.9% of the FP8 model's zero-shot accuracy; the same next token 86% of the time |
| Maya-S | DeepSeek-V4.1-Flash-1017 | 178 GB, plus about 203 GB of Unsloth Engram tables | the same next token as the original 92.6% of the time |
GLM-5.3-Flash has 321 billion parameters and about 18 billion active per token. Native context reaches 1 million. The catalog profile serves 262,144, the window of the upstream PRO 6000 measurement. DeepSeek-V4.1-Flash-1017 is the 17 October 2026 weight refresh, 748 billion parameters. The profile serves 131,072, matching the vLLM path already in the catalog. Native context stays at 1 million.
Two machines, a lower bill
| Profile | Machine | Catalog GPU band | Upstream figure |
|---|---|---|---|
glm53flash-maya-pro6000 | 1× RTX PRO 6000, at least 64 GB RAM | 1.2–2.0 USD/h | 81 tok/s decode, 3102 prefill, 256k context |
dsv41-maya-h200 | 1× H200, at least 192 GB RAM, NVMe | 2.5–5.0 USD/h | tok/s not published for a single H200 |
The 81 tok/s row is Maya's, version 1.1.7, on an RTX PRO 6000 Blackwell, a Ryzen 7 9800X3D and 61 GB of RAM, with Maya-M. SophyAI measured the Unsloth path of the same GLM-5.3-Flash on the same card class at about 51 tok/s and 1.70 USD/h. The GLM-5.3 flagship on an H200, in one of our sessions, is 3.68 USD/h. Those are different models: Flash and the flagship. The Maya profile is the Flash route, with a higher reported decode rate, on one workstation GPU.
For DeepSeek, vLLM on 8× H200 is listed at 16–32 USD/h, with the full weights at about 510 GB. One H200 sits at 2.5–5.0 USD/h. Maya-S puts hot experts on the card and reads the others from RAM or NVMe. More nearby memory means more speed. The exact figure on one H200 is one we will measure: Maya does not publish it.
API and limits
The server speaks OpenAI at http://127.0.0.1:8080/v1 and Anthropic at /v1/messages, with tool calling and images. At startup Maya pins most free RAM for the experts. The recipes are reference: no SophyAI session, no tok/s of ours, launchable=false.
The optional DFlash2 drafter is CC BY-NC-ND 4.0. The catalog profiles do not download it. The Strata profiles on a 5090 and a PRO 6000, in the other post, are the same runtime on the lower price step.