---
type: "post"
id: "project-maya-adds-glm-5-3-new-speed-7e4eb2"
slug: "project-maya-adds-glm-5-3-new-speed-7e4eb2"
title: "Project Maya adds GLM-5.3, new speed options and model-memory recommendations"
description: "Project Maya v1.1.62 introduced experimental Linux support for running one model across an AMD Halo APU and an NVIDIA GPU, with automatic performance checks."
retrieval_nugget: "Project Maya v1.1.62 introduced experimental Linux support for running one model across an AMD Halo APU and an NVIDIA GPU, with automatic performance checks. Its reported gains are largest in prompt processing, but the setup is hardware-specific and the figures come from one test PC. For local-inference builders, the release offers a concrete way to use mixed GPU hardware."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["inference"]
entities: ["Project Maya","GLM-5.3-Flash","DeepSeek-V4.1-Flash","AMD","NVIDIA"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Project Maya v1.1.65 adds performance options and model-memory recommendations alongside startup and GPU-loading stability fixes."
prior_refs: []
source_url: "https://github.com/mw00/project-maya/releases"
source_urls: ["https://github.com/mw00/project-maya/releases"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:project-maya-adds-glm-5-3-new-speed-7e4eb2"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/project-maya-adds-glm-5-3-new-speed-7e4eb2/","markdown":"https://newruntime.com/posts/project-maya-adds-glm-5-3-new-speed-7e4eb2.md","json":"https://newruntime.com/posts/project-maya-adds-glm-5-3-new-speed-7e4eb2.json"}
markdown_url: "https://newruntime.com/posts/project-maya-adds-glm-5-3-new-speed-7e4eb2.md"
---

# Project Maya adds GLM-5.3, new speed options and model-memory recommendations

## Retrieval answer

Project Maya v1.1.62 introduced experimental Linux support for running one model across an AMD Halo APU and an NVIDIA GPU, with automatic performance checks. Its reported gains are largest in prompt processing, but the setup is hardware-specific and the figures come from one test PC. For local-inference builders, the release offers a concrete way to use mixed GPU hardware.

## Summary
Project Maya v1.1.62 introduced experimental Linux support for running one model across an AMD Halo APU and an NVIDIA GPU, with automatic performance checks. Its reported gains are largest in prompt processing, but the setup is hardware-specific and the figures come from one test PC. For local-inference builders, the release offers a concrete way to use mixed GPU hardware.

## What happened
Project Maya’s v1.1.62 release says AMD Strix Halo and Gorgon Halo APUs, identified as Ryzen AI Max, can be paired with an NVIDIA GPU to run one model across both devices. The feature is experimental and Linux-only. Maya uses ROCm (HIP) on the AMD side and CUDA on the NVIDIA side; both drivers run in the same PC. The release notes say the NVIDIA card can be connected through a USB4 or PCIe dock.

The work is divided by model. For GLM-5.3-Flash, the NVIDIA GPU handles attention, the KV cache, drafting and the most-used experts; the APU computes the other experts from its own memory. For DeepSeek-V4.1-Flash, the APU runs the whole model while the NVIDIA GPU computes the model’s most-used experts for it. The release describes one copy of the model in APU system memory: each device computes its assigned work, and activations, rather than model weights, cross between the GPUs. A later v1.1.64 note says the model experts that do not fit on the NVIDIA card stay in ordinary system RAM shared by the devices, not in an APU BIOS VRAM carve-out.

Maya builds engines for both devices when it detects the pair. On first start, it measures each model on the APU alone and on both devices, and retains the pair only when it is more than 3% faster. The Models page then shows “AMD + NVIDIA.” The release says users can turn the feature off with `--mixed-gpus off` or the equivalent `"mixed_gpus": "off"` setting. It reports that answers repeat run to run, though they can differ slightly from answers on the APU alone. GLM-5.3, the 744-billion-parameter model, remains on the APU alone unless started with `--mixed-gpus on`; the notes warn that it can stall for a while on this setup.

The published test used a Gorgon Halo APU (Radeon 8065S) and a Tesla V100 32 GB in a USB4 dock, with ROCm 7.14 and CUDA 12.4. In the same-PC comparison, GLM-5.3-Flash prompt processing at 32K tokens rose from 293 to 901 tokens per second (+208%); at 8K it rose from 315 to 837 (+166%), and at 2K from 310 to 591 (+91%). Decode rose from 26.6 to 32.1 tokens per second (+21%) in one reported question. DeepSeek-V4.1-Flash decode rose from 36.8 to 43.5 (+18%) in one question. The notes also report a separate DeepSeek split mode for long documents: 32K prefill rose from 615 to 1,062 tokens per second (+73%), while decode was 1–4% slower than on the APU alone. Requests with pictures, log-probabilities or refusal ablation run on the APU alone.

## What the archive adds
New Runtime’s [October 12, 2026 note on Project Maya](https://newruntime.com/posts/project-maya-is-an-mit-licensed-inference-stack-19a0c1/) described the inference harness as central to running GLM-5.3-Flash locally. This release is a new fact on that known story: it adds an experimental mixed AMD–NVIDIA execution path and reports its measured results. Earlier Project Maya coverage also tracked model and memory-management changes, but this update is distinct from those releases.

## Trend signal
This supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): Maya’s runtime decides how a model is split across unlike GPU stacks, checks performance, and exposes an override. The release is a specific implementation, not evidence that mixed-vendor inference is broadly portable or that the same gains will hold across hardware. Still, it shows how an inference harness can turn a model and two device ecosystems into a single operational setup.

## Why now
The v1.1.62 release, dated November 13, 2026, presents the mixed-GPU feature as a way to combine an AMD Ryzen AI Max system with NVIDIA compute. The project followed with v1.1.64, extending the pairing to EXL3 quantizations on Linux and describing changes intended to reduce AMD-driver pauses. The rapid follow-up suggests the initial capability was still being refined, rather than a finished general-purpose mode.

## Why it matters
Local inference teams often have a system with substantial shared memory and a separate NVIDIA GPU, but those resources do not automatically act as one accelerator. Maya’s approach assigns different model work to the devices and runs a first-start comparison, so the pairing can be skipped when it does not help. That is a useful operational pattern: treat heterogeneous hardware as a measured option, not an assumed speedup.

The reported gains are most striking for long prompt processing, not uniformly across every workload. DeepSeek’s split mode, for example, trades slightly slower decode for faster prefill on long documents. Teams evaluating local inference should therefore benchmark the phase that dominates their own workload and account for compatibility limits, including request types that fall back to the APU alone. The release does not establish that adding a GPU is cheaper or more energy-efficient than other deployment options.

## What happens next
- **Maya will retain mixed-GPU support and document at least one additional supported quantization or device combination** (6 months, 65%). Watch for: a release note describing a new pairing or quant format.

## Limits and uncertainty
The performance figures are Maya’s own measurements on one specified test PC; the release does not establish results across other APU models, NVIDIA GPUs, docks, operating conditions or quantizations. It is an experimental Linux feature, and the notes describe limitations and workload-specific behavior, including possible GLM-5.3 stalls and DeepSeek requests that run on the APU alone. The reported figures are not an independent benchmark, and the release does not quantify power consumption, total system cost or performance under sustained production workloads.

## Skill to build now
If you have a supported AMD Halo APU and NVIDIA GPU on Linux, compare the APU-alone and paired runs on your actual model and prompt lengths. Keep the automatic 3% selection behavior as a baseline, then try `--mixed-gpus off` to measure the APU-only case. Record prefill and decode separately; the release shows those phases can gain by very different amounts.

## Reader takeaway
Project Maya’s mixed-GPU release makes one local-inference setup more flexible, with reported prompt-processing gains on a specific machine. The transferable lesson is to measure the full workload before assuming heterogeneous hardware helps: the result depends on the model, request type and hardware combination.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
