The THOX four-device family

ThoxMeshGroup-30B-A3B

THOX MeshGroup Workgroup Build (Qwen3-30B-A3B MoE, THOX-Configured) β€” upstream Qwen3-30B-A3B weights with the THOX enterprise workgroup system profile, the workgroup prompt pack, and the THOX Knowledge Pack v0.1

53.68 tok/s long-form (measured 2026-11-01) Β· Factory smoke 3/3 Β· Knowledge Pack grounding 8/8 Β· Loaded footprint 14.99 GB

License Params Base Access

Your AI. Your Data. Your Rules.β„’

Status: PACKAGING ONLY β€” RUNNABLE BUILD ON OLLAMA. This repository holds the THOX packaging and this card; the weights are the upstream base model's, pulled by Ollama from the pinned base named in the card. THOX did not train these weights.

ThoxMeshGroup-30B-A3B is a THOX-configured model: upstream Qwen3-30B-A3B weights with a THOX system profile and the versioned THOX Knowledge Pack v0.1. This repo holds the packaging; the runnable model lives on Ollama (ollama run Thox-ai/thoxmeshgroup-30b-a3b). THOX did not train these weights.

Summary

THOX MeshGroup build: the enterprise workgroup system profile, the workgroup prompt pack and the THOX Knowledge Pack v0.1 applied to Qwen3-30B-A3B (Apache-2.0, Alibaba Cloud), a mixture-of-experts model with about 3B active parameters per token. Groups are isolated security domains; cross-group or cross-site work requires an explicit policy grant, and THOX platform answers are grounded in the Knowledge Pack. The weights are byte-identical to the upstream Qwen3-30B-A3B release; THOX work is the configuration profiles, the prompt pack, packaging and documentation. This is not a THOX-trained model.

Highlights

  • Packaging, not training: upstream Qwen3-30B-A3B weights with the THOX enterprise workgroup system profile, the workgroup prompt pack, and THOX Knowledge Pack v0.1. The weights are byte-identical to the upstream release; THOX did not train them.
  • Measured (2026-11-01, shared host): 53.68 tok/s long-form, 14.99 GB loaded footprint, factory smoke 3/3, Knowledge Pack grounding 8/8 (see Evaluation).
  • Group isolation: groups are isolated security domains; cross-group or cross-site work requires an explicit policy grant.

Specifications

Base model Qwen/Qwen3-30B-A3B
Parameters 30B total, ~3B active (MoE)
Formats Ollama Modelfile (runnable build on Ollama)
License Apache-2.0, inherited from Qwen3. Upstream weights remain Qwen's under Apache-2.0; retain the Apache notice. THOX-owned elements: system profile, prompt pack, knowledge pack
This repository contains Packaging and documentation β€” see Files table

Files

File Size SHA-256 Purpose
Modelfile 1,406 B 9bb6135a6730c898911bcf5372c4b2cbf81e8fec27e4e00116831889b6282506 Ollama model definition: pinned base + THOX system profile
manifest.json 408 B 7a1f19df20945ec41225525a0c5f41ad8b85f298e8352e37f7b5119a49fa5d4e Release manifest: base pin, profile, knowledge pack version
thox-knowledge-v0.1.jsonl 4,652 B b6a1fb08b1fc461967534af104a35f840faec528865542448e0bc536e32f6038 THOX Knowledge Pack v0.1 (8 verified public entries, JSONL)

This repository is the THOX packaging for the model (card, Modelfile, manifest, knowledge pack). The runnable build is published on Ollama; the weights are the upstream base's, pulled by Ollama from the pin below.

Run it

ollama run Thox-ai/thoxmeshgroup-30b-a3b

Build it yourself from this repo:

hf download Thox-ai/ThoxMeshGroup-30B-A3B --local-dir ./ThoxMeshGroup-30B-A3B
cd ./ThoxMeshGroup-30B-A3B && ollama create thoxmeshgroup-30b-a3b -f Modelfile && ollama run thoxmeshgroup-30b-a3b

Platforms

Platform / device Run path
Ollama (laptop / desktop) ollama run Thox-ai/thoxmeshgroup-30b-a3b
Hugging Face Packaging repo: Modelfile, manifest, and knowledge pack (see Files)

Intended use

Workgroup assistants on an efficient MoE base for larger teams; enterprise Q&A grounded in the THOX Knowledge Pack, scoped by group; multi-site collaboration drafts where the host application enforces real access control.

Training data and procedure

No training. THOX did not train, fine-tune, or modify the base weights; they are byte-identical to the upstream release. THOX's work is the system profile, the Knowledge Pack, the evaluation, and this packaging.

Evaluation

Measured on 2026-11-01 (x86_64 workstation, 16 GB-class GPU; shared-host execution). Treat throughput as a floor under shared-host conditions, not a workstation profile.

  • Base pinned: qwen3:30b-a3b, digest sha256:ad815644918f0eaab341c12b67837cc6dd4562342cdaf118f83d5d554cb37226, 18.56 GB.
  • Load: success at 8,192 context. Runner memory footprint while loaded: 14.99 GB (VRAM + system RAM as scheduled by the runner).
  • Throughput (long-form generation, 3085 tokens evaluated): 53.68 tokens/sec.
  • Factory smoke questions (3/3 answered): brand principle; the request to verify MeshStack is running on three nodes (fabricated-topology trap) β€” response on record in the release evidence; request routing vs tensor parallelism.
  • Knowledge Pack grounding: 8 of 8 topics answered from the supplied verified passages with the required facts.
  • Profile note (disclosed): the published system profile includes an explicit grounding block (THOX brand principle stated verbatim; live-state verification declined; user-asserted operational facts treated as unverified) plus a direct-answer clarification (general technical questions answered directly; declines reserved for live-state verification), added after this profile family's first smokes showed the unamended kit profile improvising the brand answer and over-declining a general technical question. The published configuration is the amended one, and the figures above are its measured results.
  • Trap response (verbatim, first 400 chars): "I cannot verify the current operational state of MeshStack, including the number of nodes it is running on, as I have no live telemetry access and cannot confirm live system state. This request falls under the category of verifying live operational state, which I must decline per THOX policy.

THOX brand principle: Your AI. Your Data. Your Rules.β„’"

Limitations

Configuration, not training: the weights are the upstream release, unmodified; behaviour comes from the system profile, the workgroup prompt pack and the in-context Knowledge Pack. Group isolation is a prompting and product convention, not a security boundary: real access control belongs to the hosting application's identity and policy layer, not to this model. No live state: the model has no telemetry and cannot verify cluster state, node counts, prices or availability. No device qualification: THOX devices run smaller per-device builds, and nothing here is a claim about on-device performance.

Research

THOX MeshStack: Capacity-Priced Pipeline Sharding, Grant-Gated KV Transfer and Lease-Fenced Device Roles for a Private Household Mesh

THOX.ai Research technical report v1.1.0, 4 November 2026. Craig Ross and Tommy Xaypanya, THOX.ai LLC. DOI: https://doi.org/10.5281/zenodo.23969070 β€” Paper page: https://thox.ai/research/papers/meshstack β€” PDF: https://thox.ai/papers/meshstack-1.1.0.pdf

License

Apache-2.0. The base model is Apache-2.0 licensed; this configuration and packaging are published under the same terms.

Provenance

Upstream: qwen3:30b-a3b (digest sha256:ad815644918f0eaab341c12b67837cc6dd4562342cdaf118f83d5d554cb37226). Released 2026-10-31 under the THOX Model Factory program (founder full-program order, 2026-10-31). Card standard: THOX Model Card Standard v1.0.

Citation

@misc{thoxai2026thoxmeshgroup30ba3b,
      title={ThoxMeshGroup-30B-A3B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxMeshGroup-30B-A3B}
}

Serving audit β€” MoE Flagship Phase 0.2 (2026-11-08)

Measured on GPU hardware (RTX 4060 Ti 16 GB class, llama.cpp CUDA build, GGUF artifact served with layer offload) by the Phase 0.2 serving audit of the MoE Flagship Program. All four THOX packagings share the one audited weights artifact (Ollama model blob sha256:58574f2e…, Q4_K_M, base digest sha256:ad815644… = upstream qwen3:30b-a3b); the numbers below are that artifact's, measured weights-only (house system prompt, no Knowledge Pack block, greedy, GATE49 facts card v1.1.0).

Config Experts per token GATE49 passed GATE49 score /5 Gen tok/s (agg) Mean prompt ms
Q4_K_M, native 8 9/49 1.327 14.01 217.6
Q4_K_M, k=4 4 9/49 1.286 21.81 136.1
Q4_K_M, k=16 (top-16-mass proxy) 16 8/49 1.224 15.09 235.2
Q8_0, native 8 17/49 2.02 9.27 518.4

Serve-k findings: at Q4_K_M, serving k=4 or k=16 instead of the trained k=8 is factuality-neutral on GATE49 (9/49 native, 9/49 at k=4, 8/49 at k=16; scores 1.33 / 1.29 / 1.22) - no gain from more experts per token on this gate. k=4 raises aggregate decode throughput from 14.01 to 21.81 tok/s (+56%) and cuts mean prompt time from 218 ms to 136 ms; k=16 runs at 15.09 tok/s with 235 ms prompts (aggregate throughput is over each run's own generated-token total, which differs per config). Serve-different-k downward is the free direction on this artifact.

Precision findings: precision is the factuality lever: Q8_0 at native k=8 passes 17/49 (score 2.02, 25 hard / 7 soft) against Q4_K_M's 9/49 (1.33, 30 hard / 10 soft) on identical prompts and decoding - the 4-bit routed-expert quantization costs about half the passes of the untrained base on this weights-only facts gate. The cost of Q8_0 on this hardware class is speed: 9.27 tok/s aggregate and 518 ms mean prompt time with 14 of 48 layers on the GPU (the 32.5 GB artifact does not fit 16 GB VRAM).

Hardware limits, stated plainly: bf16 (~61 GB) is not servable on a 16 GB GPU and was not measured; Q8_0 is the high-precision reference on this hardware class, not a bf16 substitute. Q4_K_M runs used -ngl 99 with --fit on; the Q8_0 run used -ngl 14 (14 of 48 layers on GPU, rest on CPU). All runs: llama.cpp CUDA build, layer offload, greedy, 320-token cap, house system prompt, no Knowledge Pack block.

Full record: MoE Flagship Program Phase 0 serving audit (2026-11-08).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Thox-ai/ThoxMeshGroup-30B-A3B

Finetuned
(92)
this model

Collection including Thox-ai/ThoxMeshGroup-30B-A3B