Run it
Every app is also under Use this model at the top of this page. Answers are best with the system prompt used in training:
You are an expert multimodal Strudel live-coding music AI. Reason through the music theory, rhythm math, visual mood, and synthesis parameters inside <think>...</think> tags, then output ONLY valid, playable Strudel JavaScript code.
The chat template always opens a thinking block, even when a client sends enable_thinking=false: the model was trained with thinking only. Sampling used in testing: temperature 0.6, top_p 0.95, top_k 20.
llama.cpp: then open the chat page on http://127.0.0.1:8080 and set the system prompt in its settings.
llama-server -hf omerayhan-DEV/Wirbel-Qwen3.5-2B:Q8_0 --jinja -c 8192 --temp 0.6 --top-p 0.95 --top-k 20
With downloaded files: llama-server -m Wirbel-v1-Q8_0.gguf --mmproj mmproj-Wirbel-v1-f16.gguf --jinja ....
Ollama: the system and params files in this repo set the system prompt and sampling.
ollama run hf.co/omerayhan-DEV/Wirbel-Qwen3.5-2B:Q8_0
LM Studio, Jan: search the app for omerayhan-DEV/Wirbel-Qwen3.5-2B or use Use this model, pick Q8_0 (Q4_K_M for less memory), and paste the system prompt above into the chat's system prompt.
Python, Google Colab: llama-cpp-python 0.3.36 or newer. pip install llama-cpp-python builds it from source, which takes several minutes on Colab. The notebook behind Use this model → Google Colab sends an image in its example; send text as below.
from huggingface_hub import hf_hub_downloadfrom llama_cpp import Llama repo = "omerayhan-DEV/Wirbel-Qwen3.5-2B"llm = Llama.from_pretrained(repo_id=repo, filename="Wirbel-v1-Q4_K_M.gguf", n_ctx=8192, verbose=False)system = open(hf_hub_download(repo, "system")).read()out = llm.create_chat_completion( messages=[{"role": "system", "content": system}, {"role": "user", "content": "Dark techno at 130 bpm with a rolling bassline"}], max_tokens=2048, temperature=0.6, top_p=0.95, top_k=20)print(out["choices"][0]["message"]["content"])
transformers (5.2 or newer):
from transformers import AutoModelForImageTextToText, AutoProcessor repo, version = "omerayhan-DEV/Wirbel-Qwen3.5-2B", "v1"processor = AutoProcessor.from_pretrained(repo, revision=version)model = AutoModelForImageTextToText.from_pretrained(repo, revision=version, dtype="auto")SYSTEM = "You are an expert multimodal Strudel live-coding music AI. Reason through the music theory, rhythm math, visual mood, and synthesis parameters inside <think>...</think> tags, then output ONLY valid, playable Strudel JavaScript code."messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "Dark techno at 130 bpm with a rolling bassline"}]inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt")out = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.6, top_p=0.95, top_k=20)print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
The main branch holds the latest version; older ones are tags (hf download omerayhan-DEV/Wirbel-Qwen3.5-2B --revision v1).
Evaluation
20 held-out prompts from the stage1 eval set and 5 general questions (no system prompt), answered with and without the stage1 adapter from the same seed, up to 2048 new tokens, temperature 0.6, top_p 0.95, top_k 20. Code counts as playable when it evaluates in headless Strudel and produces events over 4 cycles.
Table with columns: Without stage1, With stage1 | Without stage1 | With stage1 |
|---|
| Playable Strudel code | 0/20 | 19/20 |
| Thinking finished | 2/20 | 20/20 |
| General answers with Strudel in them | 0/5 | 0/5 |
| Average reply length (tokens) | 1824 | 321 |
Training
Each stage trains a fresh LoRA adapter (alpha = 2 × rank) on top of the merge of all earlier stages, on a Raspberry Pi 5 (16 GB, CPU only). Replay mixes earlier stages' data in against forgetting.
Table with columns: Stage, Train rows, Replay, Epochs, Steps, LoRA rank, LR, Max tokens, Eval loss, Finished| Stage | Train rows | Replay | Epochs | Steps | LoRA rank | LR | Max tokens | Eval loss | Finished |
|---|
| stage1 | 1628 | – | 1 | 407 | 16 | 0.0001 | 768 | 2.170 → 0.115 | 2026-10-31 |
Changelog
v1
First release: stage1, Strudel syntax.
- Trained on 1,628 template-generated prompt → reasoning → code examples (181 more held out for eval): full tracks, mood and scene prompts, edits to an existing pattern, bass, drums, chords, leads, euclidean polyrhythms, transforms and sample manipulation, across 10 drum-kit genres, 10 scales and modes and 124 tempos. Every example was run in headless Strudel and plays; every sound and instrument name exists in Strudel. Text only, so vision is the base model's, unchanged.
- 407 steps on a Raspberry Pi 5, about 38 hours including a power cut (resumed from a checkpoint).
- What works: replies are short and finish their thinking; the code plays in most cases; tempo math (bpm →
cpm) and keys are right; general questions get normal answers.
- Known limits: it often ignores the instrument or style you ask for (a Rhodes becomes a sawtooth, breakbeats a plain kit, swing is dropped) and leans on the RolandTR808 kit. It sometimes invents a method or writes invalid mini-notation, so check the code in strudel.cc. The eval set comes from the same templates as the training data, so the scores above measure format and validity more than creativity. Stage 2 targets prompt adherence and variety.
License
Apache-2.0, like the base model Qwen/Qwen3.5-2B by the Qwen team; its LICENSE file is included.