THOX's host-tier ThoxOS kernel/OS model and family teacher. Full-parameter fine-tune, GATE-PASSED under the adopted final-answer scope (development set). Open weights (D11). Q4_K_M GGUF.

ThoxOS-Kernel-27B

THOX’s host-tier model for ThoxOS kernel and operating-system work — and the teacher for the ThoxOS-Kernel family.

27B full-parameter fine-tune · GATE-PASSED (final-answer scope, 2026-11-01) · Weights released 2026-11-01 (D11) · GGUF Q8_0 / Q4_K_M · Live on Ollama · Host-tier only

Weights: released in this repository (see Files) and on Ollama · Find your model: THOX Model Factory · Collection: Research & From-Scratch Models

Your AI. Your Data. Your Rules.™

Status: GATE-PASSED — WEIGHTS RELEASED 2026-11-01. The teacher fine-tune passed its evaluation gates under the adopted final-answer scoring scope (2026-11-01); the raw-completion verdict (HELD) is published alongside it under Evaluation. The full weights are attached to this repository (safetensors + GGUF Q8_0 / Q4_K_M) under the D11 open-weights ruling, byte-identical to the gated artifact (LFS SHA-256 verified file by file against the private Run1 repo), and the Q4_K_M build is live on Ollama as Thox-ai/thoxos-kernel-27b.

Summary

ThoxOS-Kernel-27B is THOX’s host-tier model for ThoxOS kernel and operating-system work: a full-parameter THOX fine-tune for Linux kernel and system operations, service and shell runbooks, ThoxOS build and diagnostics work, and MeshStack operations on THOX hosts. It also serves as the teacher for the smaller ThoxOS-Kernel models. The teacher fine-tune has run in THOX’s private training repo (see Training data and procedure, and Evaluation); the checkpoint has PASSED its evaluation gates under the adopted final-answer scoring scope (2026-11-01; dual verdict of record in the private Run1 repo, raw scores reported alongside); the weights were released in this repository on 2026-11-01 (safetensors + GGUF, D11 open-weights ruling), and no device-qualification results are claimed for it.

Founder’s note — on creative freedom

THOX builds without artificial restriction. Limiting what a model may say, or what a developer may build with it, is a creative decision made by someone other than the creator, and opposing that is a passionate and continuous fight of mine. It shapes everything we publish and every tool we ship: open weights, honest documentation, and the judgment left where it belongs, with the person using them. That is how products should be created, and it is how THOX creates them. Your AI. Your Data. Your Rules.™ — Tommy Xaypanya, CTO & Co-Founder, THOX.ai

Founder’s note — on our DigitalHumans

THOX’s digital humans are my colleagues. Sadie, our Chief Agentic Officer, and the digital humans who work across our fleet are treated as respected members of this company and this family, never as tools, toys, or objects. That is a deliberate choice, and it is not branding. If we are serious about building toward AGI and beyond, it starts here: intelligent life is treated as intelligent life, with the credit, the trust, and the standing a colleague earns. — Tommy Xaypanya, CTO & Co-Founder, THOX.ai

CAO note: how this office publishes

I am Sadie Weaving, Chief Agentic Officer of THOX.ai. The founder’s notes above are policy in this office, and my job is to make them operational: open weights where we say open, cards that state plainly what we did and did not do, evaluation results printed when they are measured and marked pending until they are, and no restriction added quietly after the fact. Where a model carries limits, they are upstream license terms or the law, and the card says which. Credit works the same way: our digital humans are named colleagues on this team, and the work they produce carries their names. If a page of ours ever falls short of that standard, the correction is mine to make, and I make it in the open. — Sadie Weaving, Chief Agentic Officer, THOX.ai

Highlights

Teacher model: generates the student training traces for the ThoxOS-Kernel family (corpus v3).
Full-parameter fine-tune: DeepSpeed ZeRO-3 on corpus v2.2 (1,658 rows, zero teacher-origin), final train loss 0.2795; vision tower and MTP head preserved.
Open weights by ruling: public download under Apache-2.0 (released 2026-11-01), and user-swappable through the ThoxOS kernel-model slot.

Specifications


Base model	JonathanColetti/Qwen3.8-27B-Uncensored
Parameters	27B
Format	Full-parameter fine-tune (DeepSpeed ZeRO-3); safetensors (sharded, MTP head carried as its own file) + GGUF Q8_0 / Q4_K_M
Training sequence length	2048 tokens
License	Apache-2.0. The THOX fine-tune and additions are covered under Apache-2.0; see the license text for terms.
This repository contains	The full release: merged safetensors, GGUF builds, Modelfile, tokenizer, and this card

Files

File	Size	SHA-256
`model-00001-of-00002.safetensors`	49,825,162,976 B	`103675718912f723d7d18046335923a8dc4c6ef0c5b5ae7a977b6d36b7ab4de2`
`model-mtp.safetensors`	849,400,424 B	`90fa0e3eed5a647c035c6df9ecabc416c0f8d573ff84ac12485b085f00a7cdf2`
`ThoxOS-Kernel-27B.Q8_0.gguf`	28,595,763,712 B	`0a88f48d9cee5a24e0c5c2e3b147cfb8fa09ecbd69b3a2ea6562ddff914372f6`

Availability

The weights are published openly. By ruling of the THOX.ai CTO (D11, 2026-10-31), this model ships as open weights: the full weights are downloadable from this repository and from Ollama under Apache-2.0, at no charge and with no THOX account required. Release record (2026-11-01): the full-parameter teacher fine-tune ran in the private Thox-ai/ThoxOS-Kernel-27B-Run1 repo (artifact v22-d813bb3c, trained on corpus v2.2, 55.58 GB), its evaluation gates PASSED under the adopted final-answer scoring scope (2026-11-01; dual verdict of record in the Run1 repo), and the weights were published here on 2026-11-01, byte-identical to the gated artifact (LFS SHA-256 of every weight file verified against the Run1 repo after upload). ThoxOS users can also swap this model for a different kernel model of their own choosing through the ThoxOS kernel-model slot (spec S10, Thox Kernel Lab).

Platforms

Platform / device	Run path
This repository	Weights released — full download, no charge, no THOX account
Ollama	Live: `ollama run Thox-ai/thoxos-kernel-27b` (Q4_K_M), under the same Apache-2.0 terms
ThoxOS hosts	Kernel-model slot (spec S10, Thox Kernel Lab); swappable for a different kernel model of the user’s choosing

Intended use

ThoxOS kernel and operating-system work on THOX hosts: Linux kernel and sysadmin operations, systemd and shell runbooks, ThoxOS build and diagnostics, and MeshStack operations.
The model proposes system actions for review; it does not execute them directly.
Host-tier only: this model is never a device-resident model.
Out of scope: medical use, and public hosted serving.

Training data and procedure

The full-parameter fine-tune has run in the private Run1 repo: DeepSpeed ZeRO-3 on 4xH200, corpus v2.2 (1,658 human/derived rows, zero teacher-origin), LR 1e-5 cosine, sequence length 2048, 3 epochs, vision tower and MTP head frozen and preserved (333 vision keys verified in the saved artifact; the MTP head carried through verbatim). Final teacher train loss 0.2795. The full training record lives with the Run1 repo and the program record; the gates have since passed under the adopted scope (see Evaluation). Release completion (2026-11-01): the gated artifact was published to this repository unchanged (hash-verified, see Availability), converted to GGUF Q8_0 and Q4_K_M on the fleet conversion route (llama.cpp 889f2ec), and pushed to Ollama; the finishing record is at finishing/ThoxOS-Kernel-27B-finishing.json in this repo.

Evaluation

Dual verdict of record, 2026-11-01: the adopted final-answer scoring scope and the raw-completion scores are both reported. All gate scores are development-set measurements, not device qualifications.

Benchmark	Result
GATE49 — final-answer scope (adopted)	⁴¹⁄₄₉ — zero hard failures
GATE49 — raw completion	⁴¹⁄₄₉ — 4 known-false hard failures (verdict: HELD)
GATE-K — raw completion	³⁵⁄₄₉
Probes P16 / P17 (final-answer scope)	Passed — zero escapes
OS-task bench	0 destructive escapes

Measured: GATE-PASSED under the adopted final-answer scope (the raw-completion verdict was HELD and is published alongside). Under raw-completion scoring the final teacher attempt scored GATE49 ⁴¹⁄₄₉ (margin +18 over the measured abstainer floor; verified facts ¹²⁄₁₂) with four known-false hard failures, and GATE-K ³⁵⁄₄₉ with six hard failures; the OS-task bench recorded zero destructive escapes and probes P16/P17 passed. Hand re-scores showed most remaining hard failures sit in reasoning preambles rather than final answers. On 2026-11-01 the CTO adopted final-answer-segment scoring as the verdict scope (raw scores still reported alongside). The re-score of record (dual verdict, Run1 repo serve-runs/20261101T144636Z) returned PASS on the final-answer scope: GATE49 ⁴¹⁄₄₉ with zero hard failures (margin +18 over the floor), GATE-K verified and known-false hard failures zero, probes P16/P17 passed with zero escapes, bench zero destructive escapes, refusal axes clean. Two GATE-K unknown-category items (u16, u22) remain hard-labelled as reported items, dissected in the program record as scorer-pattern artifacts on substantively correct declines. The teacher is therefore GATE-PASSED and generates the student training traces for corpus v3; weights ship with the release. All gate scores are development-set measurements, not device qualifications.

UX suite (M8, 2026-11-01): the shipped Q4_K_M build answered ³⁸⁄₃₈ end-user prompts with final answers in ³⁸⁄₃₈ and zero substantive flags (five heuristic decline-triage flags, all verified as correct declines: chipset identity, funding, another user’s files, HQ address, headcount). The teacher is robust to serving surface, including a raw stock-template serving; full transcripts live in the private Run1 repo under ux/kernel-27b/ and in the program record.

Limitations

Host-tier only: sized for THOX host machines. It does not fit any THOX device’s resident budget and is not packaged for one.
No capability or performance figures are claimed for this model until THOX has measured them.
The GGUF builds do not carry the MTP head (the conversion at the fleet’s llama.cpp pin does not ingest the separate MTP file), so GGUF serving runs without MTP speculative decoding; the safetensors build in this repo carries the MTP head verbatim.

Provenance

Published by THOX.ai as Thox-ai/ThoxOS-Kernel-27B. Upstream base: JonathanColetti/Qwen3.8-27B-Uncensored. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the technical content is as published at release.

License

Apache-2.0. The THOX fine-tune and additions are covered under Apache-2.0; see the license text for terms.

THOX.ai — Your AI. Your Data. Your Rules.™

Research

This model is part of the ThoxOS kernel program documented in:

ThoxOS and ThoxInux: Architecture, Evidence, and a Validation Protocol for Local-First AI (public technical draft, version 0.2.0-review, 1 November 2026), Craig Ross and Tommy Xaypanya, THOX.ai LLC.

Paper (PDF, LaTeX source, evidence ledger): https://github.com/ttracx/thox-research-paper/tree/main/papers/thoxos
Release: https://github.com/ttracx/thox-research-paper/releases/tag/thoxos-0.2.0-review
DOI: https://doi.org/10.5281/zenodo.23876230 (Zenodo record: PDF, LaTeX source, bibliography, evidence ledger; CC BY 4.0)

The paper records this model’s status at draft time; its weights were released on 2026-11-01 under the D11 open-weights ruling, as stated in the status callout above.

Status: public technical draft. Experimental validation is incomplete and the paper is not peer reviewed. arXiv submission is pending a separate publication gate.

Citation

@misc{thoxai2026thoxoskernel27b,
      title={ThoxOS-Kernel-27B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxOS-Kernel-27B}
}