Parable· agent-trace fine-tune

Nanbeige/Nanbeige4.2-3B, fine-tuned on Claude Fable 5 agent traces

This is not Nanbeige's Nanbeige4.2-3B. It is that model with a QLoRA adapter trained on genuine agent session traces, and, as far as we can find, the first fine-tune of this base published anywhere.

Parable

AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5

4.17B parameters (3B non-embedding)Apache-2.0

Training 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.

Nanbeige4.2-3B with a QLoRA adapter trained on genuine Claude Fable 5 agent sessions: the chat variant of the recipe, which keeps the <think> reasoning block and prose answers and drops tool-call targets. The base was released on 2026-07-24; this fine-tune shipped five days later and is, as far as we can find, the first published fine-tune of it anywhere.

The architecture is unusual: 22 transformer layers run twice per forward pass (num_loops = 2), giving an effective depth of 44 in a 4.17B-parameter footprint, with a 262k-token context window. That loop count is easy to lose in a merge-and-convert pipeline (the GGUF converter silently defaults it to 1 if the config drops the field, which would ship a half-depth model), so this pipeline asserts num_loops = 2 at the merge, at conversion, and in the final GGUF header.

What is measured: held-out trace loss falls from 2.333 to 2.133 on sessions the model never saw during training: an 8.6% reduction, the smallest in the series, against the base that starts closest to the trace distribution. Training ran at a 3,072-token window chosen from a measured retention sweep: at 2,048 tokens only 24% of traces kept their answer span; at 3,072, 98% did. Training checkpoints were relayed to Hugging Face every 50 steps, letting the run move between GPUs (it started on a P100 and finished on an A10G at 4.6x the step rate) without losing progress.

What is not measured: no public agent benchmark has been run on this checkpoint. The base's own card reports strong agentic scores (SWE-bench Verified 63.6, Terminal-Bench 2.0 44.1; vendor self-reported, not reproduced here). This tune trains the reasoning voice, not tool-calling, and prose-only SFT should be expected to trade some benchmark sharpness for style fidelity. If you want maximum harness scores, use the base; if you want its capability with a transparent reasoning trace in local use, this is the variant.

On benchmarks

There is no benchmark table on this page, deliberately. Nothing public has been run on this checkpoint yet, and the base's own numbers are vendor self-reported. Quoting them next to this model would imply a comparison nobody has measured. 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	2.57 GB	Recommended
Q5_K_M	2.99 GB	Higher quality
Q6_K	3.42 GB	Near-lossless
Q8_0	4.43 GB	Maximum quality
F16	8.34 GB	Unquantised

All quants: AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5-GGUF →

Run it

llama.cpp (build from 2026-07-27 or newer)

# GPU offload recommended: the arch's CPU path in llama.cpp is
# still maturing (days old) and hangs in our CPU-only tests.
llama-cli -m Parable-Nanbeige4.2-3B-Claude-Fable-5-Q4_K_M.gguf \
  -ngl 99 -c 4096 \
  -p "Plan the steps to debug a failing CI job." --temp 0.6

Transformers (full precision)

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained(
    "AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5",
    torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)

# use_cache=False needed on transformers >= 4.47: the base's modeling
# code predates a cache-API removal and crashes otherwise
out = model.generate(ids, max_new_tokens=512, use_cache=False)

Try others

Where to get it

Hugging FaceFull-precision weights, the model card and every file.

GGUF buildsQuantised for llama.cpp and LM Studio. Sizes in the table above.

Base model: Nanbeige/Nanbeige4.2-3B
Training data: Glint-Research/Fable-5-traces and Roman1111111/gpt5.5-terminal