Qwen3-30B-A3B-derive-sens-q4_0-q3_k.gguf is a 17.1 GB GGUF of Qwen/Qwen3-30B-A3B, made from Qwen3-30B-A3B-Q8_0.gguf by runner derive, for local inference in GGUF engines. It is not the publisher's model, only measured against it, and it was measured on one text only: the project's gold corpus (gold-corpus-2k.txt: English technical documentation with code, English prose and one Swedish passage), 2000 positions. No task benchmark was run.
Weakest number first: top-1 agreement with the publisher's float32 model is 92.90%: at about one position in 14 the file's most likely next token differs from the reference's.
This file passes the bar: mean KLD 0.0290 against the publisher's float32 model (bar 0.05) and margin-qualified top-1 99.01% (bar 97%).
Against other files in the same window
Each file measured against the same reference at the same positions; paired tests against this file. Files closer than this one come first.
- Our hand-made selective-plan file
Qwen3-30B-A3B-expq4_0-attnq8_0.gguffrom Joakimpalm-Zen/Qwen3-30B-A3B-selective-attnQ8_0-expQ4_0-GGUF (17.99 GB, 0.88 GB larger than this file) is closer to the float32 model than this file on mean KLD: 0.0279 against 0.0290 (paired Wilcoxon p 0.02). Margin-qualified top-1: 98.82% against this file's 99.01% (13 positions only this file got, 10 only that file got; McNemar p 0.68, a tie). - This file is closer to the float32 model than Qwen's official
Qwen3-30B-A3B-Q4_K_M.gguffrom Qwen/Qwen3-30B-A3B-GGUF (18.56 GB, 1.45 GB larger than this file) on mean KLD: 0.0290 against 0.0909 (paired Wilcoxon p 1.9e-235). Margin-qualified top-1: 94.80% against this file's 99.01% (73 positions only this file got, 9 only that file got; McNemar p 1.4e-13, a difference). That file fails the bar.
Size
| bytes | |
|---|---|
parent Qwen3-30B-A3B-Q8_0.gguf |
32483931648 |
| this file | 17106040320 (52.7% of the parent) |
The recipe's target was 17106040320 bytes: 11 groups demoted from q4_0 to q3_k, the ones that add the least round-trip error per byte saved.
Recipe
runner derive -m Qwen3-30B-A3B-Q8_0.gguf -o Qwen3-30B-A3B-derive-sens-q4_0-q3_k.gguf --imatrix Qwen_Qwen3-30B-A3B.imatrix --target-size 17106040320 --high q4_0 --low q3_k --match _exps --license apache-2.0 --corpus gold-corpus-2k.txt --window 8:2:2000 --reference hf:Qwen/Qwen3-30B-A3B@ad44e777bcd18fa416d9da3bd8f70d33ebb85d39 --reference-arg --upcast-at-use
Fidelity
| reference | positions | mean KLD | median KLD | max KLD | top-1 | margin-qualified top-1 |
|---|---|---|---|---|---|---|
| publisher, float32 | 2000 | 0.0290 | 0.0134 | 1.1791 | 92.90% | 99.01% of 1519 |
KL is over the reference's 20 most likely tokens, both sides renormalised; a position is margin-qualified when the reference's top two are more than 0.5 nats apart. Positions 8 + n x 2 for n < 2000 of gold-corpus-2k.txt.
Reference
The publisher's model Qwen/Qwen3-30B-A3B at ad44e777bcd18fa416d9da3bd8f70d33ebb85d39 (Qwen3MoeForCausalLM), run in float32 arithmetic, BF16 weights upcast at use, by scripts/gold-logits.py. The bar (mean KLD at most 0.05, margin-qualified top-1 at least 97%) is stated against this reference only.
License
apache-2.0, given with --license: the parent file declares no general.license.
Credits
- Importance matrix: bartowski's
Qwen_Qwen3-30B-A3B.imatrixfrom bartowski/Qwen_Qwen3-30B-A3B-GGUF at46f17e07(Apache-2.0), computed on his calibration dataset, with thanks from him to kalomaze and Dampf for its creation. - Parent: Qwen's own
Qwen3-30B-A3B-Q8_0.gguffrom Qwen/Qwen3-30B-A3B-GGUF.
Records
Qwen3-30B-A3B-derive-sens-q4_0-q3_k.gguf.derive.json holds both files' sha256, the recipe, the window and the measurement; Qwen3-30B-A3B-derive-sens-q4_0-q3_k.gguf.quant.json is the rewrite's own record.