---
license: apache-2.0
language:
- en
base_model:
- LiquidAI/LFM2-8B-A1B
pipeline_tag: text-generation
library_name: transformers
tags:
- distill
- finetune
- unsloth
- moe
- mixture of experts
---

<h2>LFM2-8B-A1B-Instruct-Quantum-IQ1C-Qwen3.6-35B-A3B-DISTILL2</h2>

A LFM 8B-1A8 (MOE, 32 experts) trained on a custom built DISTILL dataset using Qwen 3.6 35B-A3B (MOE, 256 experts, in thinking mode, with 24k context) as a teacher
via LMSTUDIO "server" (using custom python scripts) using IQ4_XS NEO-CODE-Dimatrix (Imatrix) custom quant.

Model training via Unsloth on local hardware via custom training scripts.

Training notes:
- Model intelligence has jumped.
- Serious improvements in formatting, style, and prose.
- Rep pen of 1, (off) can now be used (whereas the root model required 1.05 to 1.1).
- Model maintained as "instruct" rather than converting to "thinking".

This is an INSTRUCT model that runs at over 500 t/s on a 5090 GPU, 50-100 t/s on CPU.

<B>SETTINGS/MODEL INFO:</B>
- 128k max context, suggest 8k min context window.
- Rep pen 1. [off] ; note model maker suggests: rep pen 1.05
- Temp .2 to 1.2 ; temp of 1 used during testing.
- 4 of 32 experts activated by default.
- Model is moderately uncensored.
- Suggest quant IQXxs (imatrix) or Q5/Q6 non imatrix ggufs.

```
IN HOUSE BENCHMARKS [by Nightmedia]:

         arc-c arc/e boolq hswag obkqa piqa  wino

LFM2-8B-A1B-Instruct-Quantum-IQ1C-Qwen3.6-35B-A3B-DISTILL2
q8-hi    0.484,0.702,0.796,0.649,0.414,0.755,0.616

LFM2-8B-A1B-Instruct-Quantum-IQ1C-Qwen3.6-35B-A3B-DISTILL
q8-hi    0.471,0.677,0.783,0.647,0.390,0.740,0.616

---

BASE UNTUNED MODEL:

LFM2-8B-A1B
mxfp8    0.460,0.575,0.829,0.624,0.394,0.711,0.567
```

---

<H2>Example Generations:</H2>

Q6_K, rep pen: 1 [off], temp:1 non imatrix

NOTE: Some formatting may be lost.

---

EXAMPLE #1

---

[coming soon]