Lobo Qwen3.8-27B GSQ-RCO IQ3_S + MTP GGUF

This repository contains the single-file deployment model for Lobo, an SM120-specialized Qwen3.8-27B inference appliance for one RTX 5070 Ti 16 GB.

Artifact

This is one complete GGUF at runtime. Users do not need the source GGUFs below unless they want to reproduce the assembly.

Provenance

The artifact is a deterministic payload-preserving assembly of public Apache-2.0 inputs:

• Base model: Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
• Base quantization: ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF, revision d562806dbafae37109975e970aae91b43e73b440, file Qwen3.8-27B-GSQ-RCO-IQ3_S.gguf.
• MTP donor: unsloth/Qwen3.8-27B-GGUF, revision f975863083b62f54a5e6fac11671c750c2bbc59c, file Qwen3.8-27B-UD-IQ3_XXS.gguf.

All 851 base tensor payloads are byte-identical to the pinned GSQ source. The 15 blk.64.* MTP tensor payloads are byte-identical to the pinned donor. No further weight training or approximation was applied during assembly.

Large MTP tensors use IQ4_XS/IQ3_S; small normalization tensors remain F32. The added MTP payload is 204,412,928 bytes.

The exact assembler, input hashes, runtime source, Windows release package, launchers, and behavioural qualification are in the Lobo public repository.

Use

Download this GGUF into the Lobo release's models directory, verify it, and start the Balanced profile:

.\tools\verify-model.ps1 -Profile Balanced
.\launchers\run-balanced-230k.ps1

For an effectively headless RTX 5070 Ti with near-zero desktop GPU residency:

.\launchers\run-mtp-headless-262k.ps1

The qualified target is Windows 11, CUDA 13.3, and NVIDIA SM120. Other platforms are not claimed.

License and attribution

The source model repositories identify their weights as Apache-2.0. This repository preserves upstream attribution and links to the exact source revisions. Lobo runtime code is distributed separately under MIT-compatible terms; see its THIRD_PARTY_NOTICES.md.