Qwen3-0.6B Q8_0 — experimental Skippy package v2

Experimental packaging-path fixture, not runtime-certified. Not included in the production Mesh LLM catalog. Requires a package-v2-capable host and compatible native runtime; old v1 releases are not supported.

This package preserves the original Q8_0 GGUF bytes and adds a source-complete tensor catalog. It is not a requantization and does not contain the old per-layer GGUF layout. Stage selection is performed by the v2 runtime.

Provenance

• Source: Qwen/Qwen3-0.6B-GGUF
• Immutable source revision: 23749fefcc72300e3a2ad315e1317431b06b590a
• Source file: Qwen3-0.6B-Q8_0.gguf
• Source SHA-256: 9465e63a22add5354d9bb4b99e90117043c7124007664907259bd16d043bb031
• Generator source commit: 10971f6d954cf0584854556c9f1cdb0edb8502c3
• Generator binary SHA-256: 1577aba2e11d7e2ab7253dc8e56b5f1b0cbe8e7bda2720c10de5c9e9e43a1a63
• Manifest native ABI requirement: 0.1.52
• Generator version: 0.76.0-rc9 (version alone does not identify compatible branch builds).

Verification

The local writer output passed verify-package-v2 against the independent original: 1 source file, 1 artifact, 310 tensors, no projectors. The original's SHA-256 was independently matched to the Hugging Face source revision's LFS metadata. This proves byte-preserving source completeness, not split inference correctness.

The package tool was reused from the recorded build above and its hash checked against its runtime manifest. The package-tool source directories and Rust ABI constants are unchanged between this generator commit and inspected PR #1672 head b9450e1ec78508968b1caf9877612232aea9961e; this is not a claim that the older host has the later HF-serving fixes.

Trial usage

Pin this repository to an immutable revision when downloading. A manual snapshot download followed by use of the local package directory is the conservative trial path. Direct hf:// split serving and two-node split/unsplit parity for this published artifact have not yet been verified. Do not interpret the old v1 preflight command's rejection as v2 validation.

Verify against a separate original:

skippy-model-package verify-package-v2 ./package --source ./original/Qwen3-0.6B-Q8_0.gguf

With a compatible v2 host/native runtime, the intended serving shape is:

mesh-llm serve --model ./package --split

A second participant and an appropriate private-mesh configuration are needed to exercise a real split. Do not mix old and new split protocol generations.

License

Apache-2.0, inherited from the upstream Qwen model. Refer to the original model card for model behavior, limitations and usage guidance.