---
license: apache-2.0
base_model:
- ibm-granite/granite-4.1-30b
pipeline_tag: text-generation
library_name: transformers
tags:
- finetune
- unsloth
- granite-4.1
- reasoning
- thinking
---

<h2>Granite-4.1-30B-Claude-4.6-Opus-Thinking-X</h2>

Fine tune of Granite 4.1 with Unsloth on local hardware to convert model from "instruct" to full "reasoning/thinking".

Other versions below.

Context: 128k 

Suggest: Temp 1, Topp: .95, Minp .05, rep pen 1 (off) OR for creative 1.05 to 1.1, min context of 8k.

Enjoy ; 

---

TECH NOTE:

The "lm_head" was split from the "embed" to improve training and quant performance prior to training. 

For GGUF quants, this will allow you to set the "output tensor" at bf16 and get stronger performance overall.

---

BENCHMARKS by Nightmedia:

```
         arc-c arc/e boolq hswag obkqa piqa  wino

Granite-4.1-30B-Claude-4.6-Opus-Thinking-Charles-Xavier
mxfp8    0.573,0.761,0.876,...

Granite-4.1-30B-Claude-4.6-Opus-Thinking-Xavier
mxfp8    0.563,0.739,0.879,0.722,0.430,0.779,0.723

THIS MODEL:
Granite-4.1-30B-Claude-4.6-Opus-Thinking-X
qx64-hi  0.526,0.696,0.894,...

- BASE UNTUNED MODEL -

granite-4.1-30b
mxfp8    0.456,0.572,0.897,0.621,0.444,0.757,0.616

granite-4.1-30b
qx64-hi  0.462,0.582,0.896,0.642,0.448,0.769,0.600
```