ibm-granite/granite-4.1-8b, fine-tuned on Claude Fable 5 agent traces
This is not IBM's granite-4.1-8b. It is that model with a QLoRA adapter trained on genuine agent session traces.
Parable
AnkitAI/Parable-Granite-4.1-8B-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.
Granite 4.1 8B with a QLoRA adapter trained on genuine Claude Fable 5 and GPT-5.5 agent sessions. On the metric this model was gated against — held-out loss on agent sessions it never saw — it is the strongest in the series: 2.030 down to 0.617, a 70% reduction.
On the internal 34-prompt qualitative suite it scores 20 of 34 fully correct and 32 of 34 fully or partially correct. The 32 is the highest partial-credit score in the series; the 20 is not the highest fully-correct score, and the 3-prompt gap to the 8B Qwen model is inside the noise band for a 34-item suite. Neither number should be used to rank the two.
Like the 8B Qwen model, this one is still on the v1 recipe. The v2 rebuild has landed on the 3B and 4B and has not yet been rolled out here, which means the session-conditioning leakage class v2 removed on the 3B has never been re-measured on this checkpoint. The GGUF repo also ships four quants, Q4_K_M through Q8_0, and no F16 build, unlike the rest of the series.
On benchmarks
There is no benchmark table on this page, deliberately. This checkpoint has never been run through HumanEval — the 75.0/70.7 figure that appears in the internal ledger against this row is footnoted there as a measurement of the 8B Qwen model, not this one. Reprinting it here would be a fabrication with a citation attached, which is the worst kind. The 34-prompt qualitative suite is reported in prose above, with its sample-size caveat.
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 | 5.12 GB | Recommended |
| Q5_K_M | 5.97 GB | Higher quality |
| Q6_K | 6.88 GB | Near-lossless |
| Q8_0 | 8.91 GB | Maximum quality |
All quants: AnkitAI/Parable-Granite-4.1-8B-Claude-Fable-5-GGUF →
Run it
ollama run parable/granite4.1-fable:8bfrom llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="AnkitAI/Parable-Granite-4.1-8B-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/granite4.1-fable
Try others
- 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.
- Qwen/Qwen3-8B — agent-trace fine-tuneHighest fully-correct score on the internal 34-prompt qual, at 23/34 — but still on the v1 recipe, and not publicly benchmarked.
Where to get it
ollama run parable/granite4.1-fable:8b- Base model
- ibm-granite/granite-4.1-8b