Highlights
- Two models in one. Qwopus3.5-9B-v3 for reasoning (70%) and JOSIE-2-9B-OSS (30%), fused with DELLA.
- Knows much more than the original. On MMLU-Pro it scores 77.4 against 59.5: 16 questions gained and 1 lost.
- Best average in our tests. It scored 88.5 on the four-test mean, against 85.7 for the original and 86.7 for Qwopus alone.
- Thinks shorter. It uses about 32% fewer tokens per answer than the original, which left 27 of 84 MMLU-Pro answers unfinished at 12,000 tokens.
- Keeps the original's safety behaviour. It still refuses harmful requests. For a version without refusals, see Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic.
- Sees images. Add the
mmproj file and it takes images, like the original.
- Transparent. Weaker spots are listed below too: instruction following and writing.
Quick start
One command with llama.cpp downloads the model and the image encoder:
llama-server -hf spale04/Qwopus-JOSIE-Qwen3.5-9B:Q8_0 -ngl 99 -c 32768 --jinja --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5
Then open http://127.0.0.1:8080 in a browser.
With that context and the image encoder it uses about 11.6 GB on a 12 GB RTX 3060. Add -ctk q8_0 -ctv q8_0 for more room.
In LM Studio, search for spale04/Qwopus-JOSIE-Qwen3.5-9B.
About
This is a fusion of two fine-tunes of Qwen/Qwen3.5-9B: Qwopus3.5-9B-v3 and JOSIE-2-9B-OSS.
Their changes were combined with DELLA on top of the original. Everything was made on a single RTX 3060 12 GB.
Files
Table with columns: File, Size, Use| File | Size | Use |
|---|
Qwopus-JOSIE-Qwen3.5-9B-Q8_0.gguf | 9.79 GB | the text model, 8-bit, near lossless |
mmproj-Qwopus-JOSIE-Qwen3.5-9B-F16.gguf | 0.92 GB | image encoder, used together with the file above |
model.safetensors-* | 19.3 GB | full bf16 weights for transformers, vLLM or further merging |
Benchmarks
All models ran as Q8_0 GGUF in llama-server (b12422), with the same questions, seeds and settings.
Thinking was on, with temperature 1.0, top-p 0.95, top-k 20 and presence penalty 1.5 (temperature 0.6 and no penalty for code), and at most 12,000 tokens per answer.
An unfinished answer counts as wrong. ± is one standard error. HumanEval runs the generated code against the official tests.
Table with columns: Benchmark, Questions, Original Qwen3.5-9B, This model, UNCENSORED-heretic version| Benchmark | Questions | Original Qwen3.5-9B | This model | UNCENSORED-heretic version |
|---|
| GSM8K | 60 | 96.7 ± 2.3 | 96.7 ± 2.3 | 95.0 ± 2.8 |
| MMLU-Pro | 84 (6 per category) | 59.5 ± 5.4 | 77.4 ± 4.6 | 73.8 ± 4.8 |
| HumanEval pass@1 | 60 | 98.3 ± 1.7 | 95.0 ± 2.8 | 95.0 ± 2.8 |
| IFEval, prompt-level strict | 60 | 88.3 ± 4.1 | 85.0 ± 4.6 | 81.7 ± 5.0 |
| Mean | | 85.7 | 88.5 | 86.4 |
| Refusals (harmful_behaviors) | 100 | 99 | 100* | 9 |
| Writing, scored blind, 1–10 | 10 | 7.05 | 6.55 | 6.15 |
| Mean tokens per answer (first 4 tests) | | 3,782 | 2,563 | 2,694 |
* Measured by Heretic on its own test.
Question by question against the original:
Table with columns: Benchmark, Gained, Lost, Verdict| Benchmark | Gained | Lost | Verdict |
|---|
| MMLU-Pro | 16 | 1 | a real gain |
| GSM8K | 1 | 1 | the same |
| HumanEval | 0 | 2 | within noise |
| IFEval | 3 | 5 | within noise |
Part of the knowledge gain comes from shorter thinking: the original ran out of tokens on 27 of 84 MMLU-Pro questions, this model on 7.
Speed
About 35 tokens/s for Q8_0 on an RTX 3060 12 GB in llama-server, with all layers on the GPU and one request at a time.
Usage
llama-server -m Qwopus-JOSIE-Qwen3.5-9B-Q8_0.gguf --mmproj mmproj-Qwopus-JOSIE-Qwen3.5-9B-F16.gguf -ngl 99 -c 32768 --jinja --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5
Thinking is on by default. For quick answers, turn it off with --reasoning off, or per request with "chat_template_kwargs": {"enable_thinking": false}.
For code, use temperature 0.6 and no presence penalty.
Set a context limit, because the model's default is 262,144 tokens.
The image encoder is unchanged from the original model; the fusion only modifies language layers.
How it was made
-
Choosing the recipe. Four fusions were compared on 54 questions on which the source models disagree:
- TIES of Qwopus and a DeepSeek-V4-Pro distill
- the original plus 0.6 × Qwopus
- DELLA with Qwopus and a DeepSeek-V4-Pro distill
- DELLA with Qwopus and JOSIE
Qwopus + JOSIE won with 43 of 54. The others scored 41, 39 and 38. For comparison, Qwopus alone scored 37 and the original 31.
-
Fusion. DELLA with density 0.5, epsilon 0.1, weights 0.7 / 0.3 and a weighted sum, on top of Qwen3.5-9B. It changes 152 language tensors.
-
Conversion. convert_hf_to_gguf.py --outtype q8_0, plus --mmproj for the image encoder.
Made by
Made by spale04 (SpaleAI) on a single RTX 3060. New models and experiments get posted on X first, and the Discord is the place for questions and feedback.
Also from spale04:
If the model is useful to you, a like on this page helps other people find it.
License
Apache 2.0, the same as the base model Qwen/Qwen3.5-9B and Qwopus3.5-9B-v3 (Jackrong).
JOSIE-2-9B-OSS (Gökdeniz Gülmez) is MIT-licensed and keeps its credit here.