---
type: "post"
id: "fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7"
slug: "fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7"
title: "Fiore-lab Expert Atlas GLM-5.3 r3.3 serves full 755B GLM-5.3 on 4× DGX Spark rings without a switch"
description: "Fiore-lab’s Expert Atlas GLM-5.3 r3.3 is an unofficial TensorFold 1.0.2 derivative for running the full 755B GLM-5.3 model across four or eight NVIDIA DGX Spark systems."
retrieval_nugget: "Fiore-lab’s Expert Atlas GLM-5.3 r3.3 is an unofficial TensorFold 1.0.2 derivative for running the full 755B GLM-5.3 model across four or eight NVIDIA DGX Spark systems. The release adds a switchless four-node ring, containerized startup, and inference controls aimed at long-context and concurrent agent workloads, giving cluster owners a practical deployment path to evaluate."
published_at: "2026-11-06"
updated_at: "2026-11-06"
record_date: "2026-11-06"
date_kind: "published_at"
topics: ["inference","compute"]
entities: ["Fiore-lab","Expert Atlas GLM-5.3","TensorFold","GLM-5.3","NVIDIA DGX Spark"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Fiore-lab’s r3.3 adds a switchless deployment of the full 755B GLM-5.3 across four DGX Spark systems; prior coverage concerned other cluster inference or benchmarks."
prior_refs: []
source_url: "https://github.com/fiore-labo/ExpertAtlas-GLM5.3"
source_urls: ["https://github.com/fiore-labo/ExpertAtlas-GLM5.3"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7"
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/fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7/","markdown":"https://newruntime.com/posts/fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7.md","json":"https://newruntime.com/posts/fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7.json"}
markdown_url: "https://newruntime.com/posts/fiore-lab-expert-atlas-glm-5-3-r3-a2c1d7.md"
---

# Fiore-lab Expert Atlas GLM-5.3 r3.3 serves full 755B GLM-5.3 on 4× DGX Spark rings without a switch

## Retrieval answer

Fiore-lab’s Expert Atlas GLM-5.3 r3.3 is an unofficial TensorFold 1.0.2 derivative for running the full 755B GLM-5.3 model across four or eight NVIDIA DGX Spark systems. The release adds a switchless four-node ring, containerized startup, and inference controls aimed at long-context and concurrent agent workloads, giving cluster owners a practical deployment path to evaluate.

## Summary
Fiore-lab’s Expert Atlas GLM-5.3 r3.3 is an unofficial TensorFold 1.0.2 derivative for running the full 755B GLM-5.3 model across four or eight NVIDIA DGX Spark systems. The release adds a switchless four-node ring, containerized startup, and inference controls aimed at long-context and concurrent agent workloads, giving cluster owners a practical deployment path to evaluate.

## What happened
Fiore-lab’s GitHub repository describes Expert Atlas GLM-5.3 as an unofficial derivative of TensorFold 1.0.2. The project targets the full GLM-5.3 `glm_moe_dsa` model, listed as 755 billion parameters, on four or eight NVIDIA DGX Spark nodes using tensor parallelism of four or eight. The current repository commit is r3.3-102, dated November 6, 2026. The project’s earlier r2-102 release established support for four- and eight-node systems; the current release updates the engine and operational tooling.

A prominent addition is a four-node ring connection that does not require a separate network switch. The repository describes ConnectX-7 in connection with that ring arrangement. It also describes a Docker release image and a one-line startup path that writes the release to the first node, discovers the nodes, locates model weights, and starts the engine. This makes the claimed operational sequence more explicit than a repository that merely supplies source files, although the materials do not establish how reliably it works in other environments.

Its release components include a Zig implementation of the model family, CUDA kernels, RoCE decode collectives, a KV pool shared across conversations, and MTP and DFlash2 drafting components. The repository also refers to saved allocations, usage records, node temperatures, and a panel for choosing four or eight nodes. These are operational features around the model runtime, rather than changes to GLM-5.3’s weights. This is a specialized multi-node deployment, not a hosted service or a general-purpose installer for arbitrary hardware.

The r3.3 update adds several controls for inference and agent use. Reasoning Control lets an operator set a thinking budget as a share of `max_tokens`, including mid-run or end-of-run summarization through the panel or API. Byte Resume is described as handling agent prompt retokenization. A MULTI mode supports concurrent conversations and live prefill/decode GPU splits. The project also describes Expert Atlas bit allocation across experts, allowing different expert widths to be selected according to use. A health check and a think block on `/health` appear among the latest update’s stated changes.

The discovery material reports approximately 1,000 tokens per second of prefill and 40 tokens per second of decode on a coding prompt of roughly 256K tokens, using fp4 KV, with context up to 1M. Those are project-reported figures in the supplied material; no independent measurement or complete benchmark methodology is provided. The repository says weights and documentation are available through GitHub and Fiore-lab’s Hugging Face banks. No price is specified, and the supplied material does not describe a managed support or service offering.

## What the archive adds
This is a new release detail on a known model thread: [GLM 5.3 and GLM 5.3 Flash are now available in Cursor!](https://newruntime.com/posts/glm-5-3-and-glm-5-3-flash-454abe/) covered GLM-5.3 Flash as selectable within Cursor’s agent environment. This release is distinct: it concerns self-hosting the full model across DGX Spark nodes, not access to the Flash variant in a hosted coding environment.

## Trend signal
The release offers a concrete example of the [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) pattern: the project’s engineering work surrounds the model with deployment, shared KV, concurrency, health checks, and controls for agent prompting. The model is the target, but the release’s practical claims depend on its runtime, network topology, and operating layer. This is a single implementation and does not establish that this architecture is broadly dominant.

## Why it matters
For teams already operating DGX Spark clusters, the project lowers some practical barriers to testing a very large open model locally. The switchless ring and container startup are directly relevant to deployment logistics, while concurrent conversations, KV sharing, and controls over reasoning and prompt resumption address serving patterns associated with agents rather than one-off chat. The combination may be useful where a team wants to retain control over model execution and experiment with long context without relying on a hosted endpoint.

The release also makes the relevant evaluation questions clearer. A prospective operator can test not only raw generation speed, but whether shared KV behavior, concurrent sessions, and the chosen reasoning budget fit its own agent traffic. The ability to change inference settings does not itself prove a better application outcome. It does, however, give infrastructure-focused teams more parameters to inspect than a fixed hosted endpoint normally exposes.

The trade-off is a substantial hardware and operations commitment. This is not an inexpensive route simply because a separate switch can be omitted, and the supplied figures do not establish cost per completed task, reliability under load, or parity with hosted systems. Builders should treat it as an implementation to inspect and test against their own hardware and workload, not as proof that local deployment is automatically faster or cheaper.

## Limits and uncertainty
The supplied repository excerpts and discovery summary do not establish independent performance, sustained throughput, hardware cost, power use, reliability, or the workload conditions behind the reported speed figures. The 1M context ceiling is reported, but the material does not show quality or latency across that full context. It also does not specify a price, support commitment, licensing terms for every included component, or a reproducible comparison with hosted inference. The system requires four or eight DGX Spark nodes and compatible networking; applicability to other hardware is not established.

The supplied material calls Expert Atlas an unofficial TensorFold derivative. It does not establish the maturity of its maintenance process, the security properties of the deployment, or whether its documented startup sequence is suitable for production environments. The project-reported benchmark should therefore be treated as a lead for reproduction, not a purchasing or architecture decision.

## Reader takeaway
Expert Atlas r3.3 is a focused self-hosting project, not a new GLM model: it packages the full GLM-5.3 for multi-node DGX Spark deployment and adds agent-relevant serving controls. Cluster owners can inspect it as a route to experimentation, but should independently validate performance, operating requirements, and licensing before planning production use.

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.
- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute 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)