README.md

---
license: apache-2.0
tags:
- merge
- not-an-original-model
---

# Mini-Merge-35B-A3B

This is a model **merge**, not an original training run. It's a plain element-wise
weighted average of two existing open-weight checkpoints (a ~40-line safetensors script;
no mergekit, since its catalog doesn't know this architecture). No pretraining, no
post-training of our own. The exact recipe ships as `merge.yaml` next to the weights.

## Ingredients
- `nex-agi/Nex-N2-mini`
- `Qwen/Qwen3.6-35B-A3B`

Merge variant: `mini-b0.2`. Full pipeline and commands:
https://github.com/PixelML/merge-theater

## Benchmarks (same lm-eval tasks, parents vs merge)

| task | mini-b0.2 | Nex-N2-mini | Qwen3.6-35B-A3B |
|---|---|---|---|
| arc_challenge | 62.4 | **62.5** | 54.6 |
| gsm8k | **90.1** | 89.2 | 29.3 |
| hellaswag | **84.0** | 83.5 | 83.5 |
| humaneval | 66.5 | **72.0** | 61.0 |
| mbpp | 63.8 | 54.6 | **66.2** |
| winogrande | 73.3 | 73.6 | **73.7** |

Aggregate (mean over shared tasks):
- mini-b0.2: 73.36
- Nex-N2-mini: 72.59
- Qwen3.6-35B-A3B: 61.37

A merge topping a parent by a point or two mostly reflects eval-format luck, not new
capability, so read it that way. The numbers reproduce with the command below.

## Reproduce
```bash
modal run modal_app.py::sweep --weights-b 0.2,0.3,0.4
```

## License
Both parents (`nex-agi/Nex-N2-mini` and `Qwen/Qwen3.6-35B-A3B`) are Apache-2.0, so this merge is released under Apache-2.0.
The merge used a small original script, not mergekit, so no LGPL applies. This card must
travel with the weights.

## Credit
Built by Claude Code (Anthropic's agent) running autonomously: the merge, the eval
pipeline, the ratio sweep, and this card.