Qwen/Qwen3-8B, fine-tuned on Claude Fable 5 agent traces
This is not Qwen's Qwen3-8B. It is that model with a QLoRA adapter trained on genuine agent session traces, and it is the largest model in the series.
Parable
AnkitAI/Parable-Qwen3-8B-Claude-Fable-5Training data carries upstream terms: Fable-5 traces are AGPL-3.0, the terminal set is MIT, and both originate from third-party assistants whose terms may reach downstream use.
Qwen3-8B with a QLoRA adapter trained on genuine Claude Fable 5 and GPT-5.5 agent sessions. It is the largest model in the series and the one with the most headroom, and it is also the one with the least evidence behind it.
What is measured: held-out trace loss falls from 2.162 to 0.712, a 67% reduction on sessions the model never saw during training. That is a real on-distribution win and it is the gate this model was shipped against. On the internal 34-prompt qualitative suite it scores 23 of 34 fully correct and 30 of 34 fully or partially correct, the highest fully-correct score in the series — though the gap to the 8B Granite model is 3 prompts, which at this sample size is noise, not a ranking. What is not measured is the capability the model was trained for: no agent benchmark has been run on this checkpoint.
This model is still on the v1 recipe. The v2 rebuild — completion-only loss masking, replay mixing, benchmark-gated checkpoints — has landed on the 3B and 4B and has not yet been rolled out here. That matters, because v1's known pathologies are the empty-answer failure inherited from Qwen3's hybrid thinking and the session-conditioning leaks that v2 was built to remove. Neither has been measured on this checkpoint.
On benchmarks
There is no benchmark table on this page, deliberately. HumanEval and HumanEval+ have not been run publicly on this checkpoint — the figure that circulated internally was measured on a different model — and the 34-prompt qualitative suite has since been demoted to an internal smoke test, because the base model outscores every shipped fine-tune on it and it therefore measures one-shot chat coding rather than agent capability. Publishing a borrowed number here would be worse than publishing nothing. An explained empty section outperforms a wrong figure permanently.
Builds
GGUF quantisations, for llama.cpp, Ollama and LM Studio. Sizes are exact file sizes from the Hugging Face API, not estimates. Q4_K_M is the one to take unless you have a reason.
| Quant | Size | Notes |
|---|---|---|
| Q4_K_M | 5.03 GB | Recommended |
| Q5_K_M | 5.85 GB | Higher quality |
| Q6_K | 6.73 GB | Near-lossless |
| Q8_0 | 8.71 GB | Maximum quality |
| F16 | 16.39 GB | Unquantised |
All quants: AnkitAI/Parable-Qwen3-8B-Claude-Fable-5-GGUF →
Run it
ollama run parable/qwen3-fable:8bfrom llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="AnkitAI/Parable-Qwen3-8B-Claude-Fable-5-GGUF",
filename="*Q4_K_M.gguf",
n_ctx=8192,
)
out = llm.create_chat_completion(
messages=[{"role": "user", "content": "Plan the steps to debug a failing CI job."}],
max_tokens=1500,
temperature=0.3,
)
print(out["choices"][0]["message"]["content"])LM Studio: lms get parable/qwen3-fable
Try others
- Qwen/Qwen3-4B — agent-trace fine-tuneThe one that fixes Qwen3's empty-answer failure: 7 of 34 prompts returned nothing before, 0 after, at coding parity.
- ibm-granite/granite-4.1-8b — agent-trace fine-tuneLargest trace-loss reduction in the series at 70% — the strongest on-distribution result, on the v1 recipe.
Where to get it
ollama run parable/qwen3-fable:8b- Base model
- Qwen/Qwen3-8B