Ankit AglaweHuggingFaceTB/SmolLM3-3B, fine-tuned on Claude Fable 5 agent traces
This is not Hugging Face's SmolLM3-3B. It is that model with a QLoRA adapter trained on genuine agent session traces, tuned end-to-end on a consumer Mac.
Parable
AnkitAI/Parable-SmolLM3-3B-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.
SmolLM3-3B with a QLoRA adapter trained on genuine Claude Fable 5 agent sessions: the chat variant of the recipe, keeping the <think> reasoning block and prose answers. The whole run happened on a 16 GB M1 Mac through MLX, 16 layers adapted, 6.7 million trainable parameters, a 4,096-token window over 4,076 training rows.
The run was paused mid-training for a workday and resumed by a scheduled script in the evening. The resume surfaced a real pipeline bug: the training driver re-issued the full 1,200-iteration budget to the resumed run instead of the remainder, which would have overtrained the adapter by roughly 70%. The fix was caught at the checkpoint arithmetic, the run was relaunched for exactly the remaining iterations, and checkpoint selection then picked iteration 200 of the final segment (validation loss 1.154) over the final iterate (1.203), which is checkpoint selection doing its job.
What is measured: held-out trace loss falls from 1.889 to 1.115 on a 226-session test split the model never saw, a 41% reduction. Because training ran against a 4-bit quantized base (the 16 GB memory constraint), the F16 merge and every quant of it cannot exceed 4-bit-base quality; the model card states the same ceiling.
What is not measured: this lane trains on trace data without a replay mix, so the impact on general coding benchmarks is unmeasured for this checkpoint, and the page says so rather than borrowing numbers from siblings. SmolLM3 is a supported llama.cpp architecture, so unlike the Nanbeige release the GGUFs run in stock Homebrew, Ollama, and LM Studio builds today.
On benchmarks
There is no benchmark table on this page, deliberately. Nothing public has been run on this checkpoint, and the series' own BFCL results on sibling models are theirs, not this model's. The held-out trace loss above is the only number this page is entitled to.
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 | 1.92 GB | Recommended |
| Q5_K_M | 2.21 GB | Higher quality |
| Q6_K | 2.53 GB | Near-lossless |
| Q8_0 | 3.28 GB | Maximum quality |
| F16 | 6.16 GB | Unquantised |
All quants: AnkitAI/Parable-SmolLM3-3B-Claude-Fable-5-GGUF →
Run it
llama-cli -m Parable-SmolLM3-3B-Claude-Fable-5-GGUF-Q4_K_M.gguf \
-c 4096 -p "Plan the steps to debug a failing CI job." --temp 0.6from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained(
"AnkitAI/Parable-SmolLM3-3B-Claude-Fable-5")
model = AutoModelForCausalLM.from_pretrained(
"AnkitAI/Parable-SmolLM3-3B-Claude-Fable-5",
torch_dtype="auto", device_map="auto")
# Output opens with a <think>...</think> block; strip it before
# showing text to end users.
out = model.generate(ids, max_new_tokens=512)Try others
- Nanbeige/Nanbeige4.2-3B — agent-trace fine-tuneNewest model in the series, on the newest base: Nanbeige4.2-3B was five days old when this shipped, and this is its first published fine-tune.
- ibm-granite/granite-4.1-3b — agent-trace fine-tuneThe density pick: 70.1 HumanEval in a 2.10 GB Q4_K_M, and a strict-qual score up from ~18/34 to 25/34 on the v2 recipe.
Where to get it
- Base model
- HuggingFaceTB/SmolLM3-3B