2026-10-20
Strata in the catalog: GLM-5.3-Flash and DeepSeek on one GPU
The Strata engine keeps a MoE model's experts across the GPU, RAM and SSD. SophyAI's lowest-cost profiles for GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 use that design, running on Project Maya.
Strata is an MIT-licensed open-source inference engine built to run a Mixture-of-Experts model on a PC. The public app serves Qwen3.8-Flash-Next. For GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 the repository to clone is Project Maya, which rewrote that engine for these two architectures. SophyAI adds two Strata profiles. They are visible and not launchable yet.
How the weights sit in memory
A MoE of this size does not fit in a 32 GB or 96 GB card. Strata, and Maya after it, splits the work into three tiers:
- the GPU holds the experts the conversation uses most;
- RAM holds the next tier;
- the NVMe holds the rest and is read in place.
Each token activates only a handful of experts. The others can stay off the VRAM. It is the same idea as the Unsloth path with --n-cpu-moe, with a runtime written for that exchange.
The two low-cost profiles
The prices below are the typical_usd_h bands already stored in gpus.json. They are not measurements from a SophyAI session. The tok/s figures are upstream Maya numbers, collected with ./maya.sh --bench.
| Profile | Machine | Weights | GPU band | Upstream speed |
|---|---|---|---|---|
glm53flash-strata-5090 | 1× RTX 5090, 96 GB RAM | Maya-M, 116 GB | 0.5–1.3 USD/h | over 30 tok/s decode, about 2000 prefill |
dsv41-strata-pro6000 | 1× RTX PRO 6000, at least 192 GB RAM, NVMe | Maya-S, 178 GB, plus about 203 GB of Engram tables | 1.2–2.0 USD/h | not published for this machine |
The GLM in these profiles is GLM-5.3-Flash: 321 billion parameters, about 18 billion active per token, native context up to 1 million. The GLM-5.3 NVFP4 flagship stays on FreeToken. The DeepSeek checkpoint is DeepSeek-V4.1-Flash-1017, published on 17 October 2026.
On the 5090 the in-catalog comparison is the Unsloth path with experts on the CPU, measured by SophyAI at about 5.6 tok/s. Maya-M's upstream figure, over 30 tok/s on the same card class, is why the profile exists. It still has to be checked in a session of ours.
For DeepSeek the cost comparison is vLLM on 8× H200, listed at 16–32 USD/h. A PRO 6000 sits at 1.2–2.0 USD/h. Maya-S reports 92.6% of the same next tokens as the original model. The 8-GPU vLLM path remains the one for the full MXFP4/MXFP8 weights, about 510 GB.
What does not start yet
The recipes are launch=reference and launchable=false. The expected API is on port 8080, OpenAI- and Anthropic-compatible, with images and tool calling. The DFlash2 drafter is non-commercial and is not part of these profiles.
strata launch from the Niko1221 repository still serves Qwen3.8-Flash-Next. For these two recipes the clone is Project Maya.