license: apache-2.0
base_model:
- Qwen/Qwen3.5-9B
- HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
language: - en
tags: - gguf
- qwen3.5
- llama.cpp
- quantization
- vocabulary-pruned
- tokenizer
- uncensored
- amd
- vulkan
- rdna1
- local-llm
Monoglot 9B: Qwen3.5-9B with its vocabulary cut to what English actually uses
One language, on purpose. Second entry in the Monoglot series, afterQwen3.8-27B-Q3-DOWN-XS-Monoglot.
Same surgery, different result: on this file the vocabulary was a quarter of the whole model, so the win
came out as raw decode bandwidth rather than as a threshold crossing.
Source file is the Q4_K_M GGUF of HauhauCS's uncensored 9B. 118,300 of its 248,320 vocabulary rows are
gone. Every surviving weight row is bit-identical to the source; nothing was retrained or re-quantised.
Credit, in order of who did what
- The idea and the first public build: fakezeta, whose
Condensed-P2release is what
showed this was worth doing. - The tool:
bsaleh03/ASCII-Condensed-prune-tools.
The policies, keep-set construction and in-quantised-space row gather are his work. That repository has
no licence, so none of its code is redistributed here; it is linked as the credit it is. - @TeksEdge for the post that surfaced it.
- The model: HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
on base Qwen/Qwen3.5-9B. The uncensored tuning is theirs; a
vocabulary prune does not add to or subtract from it, and the identical-greedy check below shows the
surviving weights produce the same text.
Mine is the application to this file, the measurements, and the parts that did not work.
What was cut, and why it is worth more here than on the 27B
Policy P1G: ASCII, plus math and typography symbols, Greek, box-drawing and status glyphs, units and
currency. The accented Latin range is deliberately not kept (that would be P2).
| vocab rows | vocab bytes | share of file | |
|---|---|---|---|
| Qwen3.5-9B source | 248,320 | 1,406 MB | 25.0% |
| This file | 130,020 | 737 MB | 14.9% |
| Tensor | Type | Before | After |
|---|---|---|---|
output.weight |
Q6_K | 834 MB | 437 MB |
token_embd.weight |
Q4_K | 572 MB | 300 MB |
| File | Q4_K_M | 5,627,044,224 B | 4,950,668,096 B |
676 MB smaller, 12.0%. A quarter of this model was vocabulary, which is the highest share of any GGUF I
have measured here, because the output head is Q6_K and the embedding is Q4_K while the body is Q4_K_M.
What it bought, measured
llama.cpp 3ba1ce0 Vulkan, RX 5700 XT (8 GB, RDNA1), flags matching the machine's production profile for
this model: -ngl 99 -ctk q4_0 -ctv q4_0 -fa on -t 8 --cpu-mask 0x3f, measured at depth 8192 (an
in-context agent turn, not an empty one). Three reps, card empty before each run.
| llama-bench | source | this file | |
|---|---|---|---|
| pp512 @ 8192 | 431.74 ± 2.91 | 434.45 ± 3.47 | flat |
| tg128 @ 8192 | 49.09 ± 0.16 | 52.07 ± 0.16 | +6.1% |
This is a different kind of win from the 27B, and worth saying plainly. The 9B already fits: 5.24 GiB
of a 7.98 GiB card, with 2.7 GiB to spare and no spill. So there was no threshold to cross, and prefill does
not move. What moves is decode, because decode reads the output head once per token and the head was 834 MB
of the 5.6 GB that had to cross the memory bus every step. Removing 397 MB of it is a 6.1% decode gain.
A naive bandwidth model predicted about 12% from the traffic reduction and reality gave 6.1%, so decode here
is not purely bandwidth-bound. That is the measured number; the prediction was wrong and I am not going to
quote the model instead.
What it costs
Token counts for the same string on both files:
| Text | Source | This file | Ratio |
|---|---|---|---|
| English prose | 30 | 30 | 1.00x |
| Python | 41 | 41 | 1.00x |
Math with ∀ ∑ π ≥ ✓ |
27 | 27 | 1.00x |
| French, unaccented | 17 | 17 | 1.00x |
| French, accented | 18 | 31 | 1.72x |
| Japanese | 15 | 40 | 2.67x |
| Chinese | 15 | 50 | 3.33x |
| Russian | 17 | 93 | 5.47x |
Note the French rows: plain ASCII French is unaffected, and it is the accents that fall back to bytes.
Correctness checks that did pass
verify_prune.py --policy P1G: all 425 non-vocabulary tensors byte-identical, gathered rows match their
source rows, keep-set exactly the policy, every metadata key preserved.- Greedy 64-token continuations byte-identical to the source for an English prompt and a code prompt.
- A
toolsrequest on both files returnsfinish_reason: tool_calls,read_file, and{"path":"/etc/hostname"}. Chat template and tool calling survive the id remap.
What I did not measure
- No instruction-following benchmark. No IFEval, no scored suite. Correctness rests on the checks above,
which prove nothing is misaligned and say nothing about instruction-following quality. - No perplexity or KL, deliberately: not comparable across a vocabulary change, because the same text
tokenises differently and the logit spaces differ in size. - One machine, one build, one quant. RDNA1, Vulkan, depth 8192, KV q4_0.
- No
mmproj, text only. - Refusal behaviour was not tested. Greedy English output is identical to the source, which is all this
card claims; it is not a moderation statement either way.
Using it
llama-server -m Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M-Monoglot-P1G-130K.gguf \
-c 32768 -ngl 99 -ctk q4_0 -ctv q4_0 -fa on
Drop-in for the source file on the same port and alias. If you need multilingual output, use the source.
Files and verification
| File | Bytes | SHA-256 |
|---|---|---|
Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M-Monoglot-P1G-130K.gguf |
4,950,668,096 | 78b2aec53b841cea479a72e11c1bf348de9a0ef53815b94d10bd2edf0ed2e8f6 |
Licence
Apache-2.0, following the base model. Weights derive from Qwen3.5-9B and HauhauCS's uncensored tuning. The
pruning tool is linked, not included, because it carries no licence.