🪶 Parable-Granite-3B v2 — trained on genuine Claude Fable 5 agent traces
A tiny local model that thinks before it answers — planning, reasoning, and terminal instincts distilled from real agent sessions.
~3 GB of RAM is all you need. Laptop, old GPU, Raspberry-Pi-class boxes with swap — the Q4 build runs
anywhere. One command and you have a private, offline reasoning model on your machine:
ollama run hf.co/AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF:Q4_K_M
The headline — v2 is a different model
v2 is a full retrain: 13× more genuine Fable 5 trace data (11,574 sessions, 16.8M tokens — corpus published) and a rebuilt recipe (completion-masked loss, replay mixing, benchmark-gated checkpoints, seed-averaged weights).
Clean answers, structured reasoning, agent instincts — and the transcript artifacts that leaked into v1's replies are gone. One trade, made on purpose: raw HumanEval-style function synthesis stays the base model's turf (81.7 vs 70.1) — v2 spends that capacity on agent behavior instead, and spends half as much as v1 did. Measurement notes below. 👇
Announcements
📌 Same links, new model. v2 replaces v1 in place — every existing Ollama command, script, and bookmark now serves v2. No migration, nothing to change.
🔮 v3 is already training. Rejection-sampled SFT: thousands of candidate solutions generated against executable tests, only verified passers enter the corpus. The goal is simple — above-base agent capability, not just clean behavior. Follow AnkitAI for the drop.
📦 Full family. This 3B is the smallest Parable. Need more headroom? 8B Granite, 8B Qwen, 4B Qwen — same recipe, no matter your hardware.
Pick your size
File
Size
Fits in
Notes
Q4_K_M
2.1 GB
~3 GB RAM/VRAM
⭐ Recommended — best size/quality balance
Q5_K_M
2.4 GB
~3.5 GB
Higher quality
Q6_K
2.8 GB
~4 GB
Near-lossless
Q8_0
3.6 GB
~5 GB
Maximum quality
F16
6.8 GB
~8 GB
Full precision, for re-quantizing
Intelligence per gigabyte: the Q4_K_M build scores 70.1 HumanEval in 2.1 GB — ~33 pts/GB; an 8B-class Q4 needs ~5 GB for its score. If RAM is your constraint, this is the family's density sweet spot.
Ollama (chat template ships inside the GGUF — zero config):
bash
1ollama run parable/granite4.1-fable:3b
2# or straight from this repo:3ollama run hf.co/AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF:Q4_K_M
llama.cpp:
bash
1llama-cli -m Parable-Granite-4.1-3B-Claude-Fable-5-GGUF-Q4_K_M.gguf --jinja \2 -p "Write a bash one-liner to find the 10 largest files in a directory tree."
LM Studio / Jan / Open WebUI: search "parable" in-app, or paste this repo URL.
Python (llama-cpp-python):
python
1from llama_cpp import Llama
23llm = Llama.from_pretrained(4 repo_id="AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF",5 filename="*Q4_K_M.gguf", n_ctx=8192,6)7out = llm.create_chat_completion(8 messages=[{"role":"user","content":"Write a Python function that retries an HTTP request with exponential backoff."}],9 max_tokens=3000, temperature=0.7,10)11print(out["choices"][0]["message"]["content"])
Thinking mode
Every answer opens with a <think>...</think> reasoning block — that's the Fable 5 heritage. llama.cpp's --jinja mode separates it automatically; strip it before showing replies to end users.
Sampling: temperature 0.7, top_p 0.95, and budget max_tokens generously (2500+) — trace-trained models think at length before answering.
Measurement notes
All numbers: identical llama.cpp harness, greedy decoding, Q4_K_M, base model measured on the same instrument. We train multiple seeds and ship the weight-average — single-run scores at 3B swing ±3 points on GPU nondeterminism alone, so most cards report their luckiest run; we ship the average and report the shipped weights' own numbers. Raw eval outputs live in this repo.
Which model should you use? Pure single-function code completion → the base model is genuinely strong there. Explanations, debugging, terminal workflows, structured reasoning, agent-style tasks → that's what Parable is trained on, and where v2 shines.
🆚 Same prompt, side by side
Real outputs, both models at Q4_K_M, temperature 0.7 — unedited except length.
Prompt:"Make this more idiomatic:"result = []; for x in items: if x.active == True: result.append(x.name.upper())
Base Granite-4.1-3B
Parable v2
Here's a more idiomatic version of the code using a list comprehension: result = [item.name.upper() for item in items if item.active] — Explanation: List comprehensions provide a concise way to create lists based on existing iterables. The new result list is constructed by iterating over each item… (continues for ~10 more lines)
result = [item.name.upper() for item in items if item.active]
Prompt:"My Python script fails with 'RecursionError: maximum recursion depth exceeded' in a JSON parser I wrote. What are the likely causes and the standard fix?"
Base Granite-4.1-3B
Parable v2
Opens with prose: "The RecursionError: maximum recursion depth exceeded error in a Python script, especially when dealing with a JSON parser, typically indicates that your recursive function is calling itself too many times without reaching a proper base case…"
Opens with a diagnosis table:Common culprits for this error — a cause / why-it-triggers / example table, then the fix. Structured like an agent working the problem, not an essay.
The pattern from real agent traces: answer first, structure over prose, no padding. (Where the base is stronger — raw single-function synthesis — is stated plainly in the measurement notes above.)
What's new in v2 (training)
The recipe follows our ongoing tech report (in preparation):
Completion-only loss masking (Hermes 3, Tülu 3) — loss on assistant tokens only, so the model learns to answer, not to imitate transcripts
30% replay mix of general instruction data (Luo et al., Biderman et al.) — the anti-forgetting lever
Session re-segmentation + sanitization — why v1 sometimes leaked agent JSON into normal chat, and v2 never does (0/34)
Benchmark-gated checkpoints (Dong et al.) instead of fixed epochs
Seed-averaged weights (model soups, Wortsman et al.) — we ship the average of multiple runs, not the lottery winner
With Claude Fable 5 now retired, genuine self-authored Fable traces are a fixed, non-renewable corpus. Unlike most models in this niche, our full training corpus is public: AnkitAI/parable-corpus-v2 — deduplicated, quality-gated, provenance-tagged.
Good to know
Fine-tuned at 2,048-token sequences; the base 128K context stays available, fine-tuned behavior is strongest in the opening turns.
Not trained for: multi-file repo navigation, vision, non-English.
Inherits Granite-4.1-3B's knowledge cutoff. Treat generated commands as drafts to review.
Evaluation
Function calling (BFCL V3, AST subset)
Measured 2026-07-29: bfcl-eval at gorilla main, prompting mode, Q4_K_M
GGUFs served by llama.cpp on a T4, base and Parable under the identical
harness. Categories: simple_python / multiple / parallel /
parallel_multiple (400/200/200/200 items). Raw generations and score
files: parable-v2-artifacts
under verify/bfcl/.
simple_python
multiple
parallel
parallel_multiple
Granite-4.1-3B base
0.848
0.790
0.710
0.665
This model (chat variant)
0.413
0.605
0.320
0.425
For tool-calling workloads, use the base model; this variant is built
for reasoning prose. The drop has a specific mechanism: sampled generations show the model
intermittently answering with args-only tool-call JSON (for example
{"base": 10, "height": 5}) instead of a function call, which the
AST scorer rejects. That is trace-scaffolding format bleeding into
standalone tasks, the failure mode the series paper names session
leakage (Section 6 of the report). The reasoning-voice strengths this variant trains for are unaffected
on prose tasks.
Base & license
Weights: Apache-2.0 (inherited from ibm-granite/granite-4.1-3b). Training data: Fable-5-traces AGPL-3.0, gpt5.5-terminal MIT — since traces originate from third-party assistants, their terms may apply to downstream training; check before commercial distillation.
The recipe, evaluation methodology and failure analysis behind this model are
documented in the tech report:
Aglawe, A. (2026). Agent-Trace Fine-Tuning of Small Language Models under
Constrained Compute. Zenodo. doi:10.5281/zenodo.21676407
bibtex
1@misc{aglawe2026agenttrace,
2 author = {Aglawe, Ankit},
3 title = {Agent-Trace Fine-Tuning of Small Language Models under Constrained Compute},
4 year = {2026},
5 publisher = {Zenodo},
6 doi = {10.5281/zenodo.21676407},
7 url = {https://doi.org/10.5281/zenodo.21676407}
8}
Acknowledgements
Glint-Research & Roman1111111 for the open trace data · IBM Granite for the base · empero-ai whose Qwable recipe inspired the series · llama.cpp
Version history
v2 (2026-07-16) — this release. 13× corpus, rebuilt recipe, seed-averaged weights, zero leakage.
v1 (2026-07) — initial release, 857-row corpus. Preserved as repo revision history.
Three gigabytes. Real Fable 5 reasoning. Yours, offline, right now.
ollama run hf.co/AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF:Q4_K_M
