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

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.

Related PixelML club work

This merge was built and evaluated on Modal, not on PixelML's GPU club hardware. It is not a member of the club-dgx-spark or club-170hx verified-hardware collections. For hardware-verified recipes, see:

• github.com/PixelML/club-dgx-spark
• github.com/PixelML/club-170hx
• PixelML/club-dgx-spark: verified on DGX Spark (GB10) (Hugging Face collection)