qwen3-8b-ralph-sub4-c1b-20260910

Compressed derivative of Qwen/Qwen3-8B submitted to the sub4 bit-tier of Bittensor subnet 40 (Ralph, model compression). Architecture and parameter count are unchanged from the parent; weights are re-stored at reduced bit-width.

License

Released under the Apache License 2.0 — see LICENSE. This is a derivative of Qwen3-8B, itself Apache-2.0; the parent's copyright and attribution notices, a description of what changed, the data used and the third-party tooling involved are preserved in NOTICE. The exact artifact is model.gguf at revision 3c21306c630dac6440b37ad67b965a4b9b787cf9 (SHA-256 b2ec80dce90258bbe9555a67983dea595d91f35cae48ca0d54c714eae8100edf); adding this model card and license files does not alter that file.

Model overview

Parent	Qwen/Qwen3-8B (Apache-2.0)
Architecture	unchanged from parent
Format	GGUF (llama.cpp)
Bit tier	sub4 (gate: ≤4.0 code bits and ≤5.0 container bits per weight)
File size	4,614,304,896 bytes
Measured bits/weight	4.0 code bits, 4.5 container bits (validator-measured)
Method	Post-training weight re-storage of the parent's own weights: llama.cpp pure Q4_K quantization with an importance-matrix (imatrix) calibration computed on the bf16 parent at 2048-token context over an exam-proportional sample of public trajectory datasets. No retraining, distillation or architectural change.

Upstream terms to preserve: Qwen3-8B's Apache-2.0 license and attribution (kept in NOTICE); llama.cpp's MIT license for the GGUF tooling; the dataset licenses listed in NOTICE (Apache-2.0, MIT, CC-BY-4.0 — attribution only).