merge-theater

A model merge done in the open, with the receipts and the same evals run on the parents, so you can see exactly how little "training" is involved.

It started with nex-agi/Nex-N2#4: a 397B model shipped as original work that was allegedly a 0.6·Nex-N2-Pro + 0.4·Qwen3.5 element-wise blend. The tell was hard to argue with. All three are the identical 397B-A17B shape, the deployed model called itself "Nex" 79% of the time, and every tensor sat thousands of std-devs off any real finetune. After the callout, the card was edited to admit a merge plus distillation, and to admit they'd uploaded the bare merge by mistake.

So this repo does the honest version: merge two models, eval the parents and the child on the same tasks, print the table, and say plainly what it does and doesn't prove.

What shipped

PixelML/Mini-Merge-35B-A3B — a weighted average of Nex-N2-mini (0.8) and Qwen3.6-35B-A3B (0.2) that beats both parents on aggregate over six benchmarks, with zero training:

	average	vs Nex-mini	vs Qwen3.6
Mini-Merge (0.8/0.2)	73.36	+0.77	+11.99
Nex-N2-mini	72.59
Qwen3.6-35B	61.37

Apache-2.0 (both parents are), openly labeled a merge.

The facts that shape what's possible

A merge blends tensors position by position, so both checkpoints must share the same base: layer count, hidden size, heads, MoE layout, tokenizer, vocab. scripts/preflight.py checks this from config.json alone and refuses otherwise.

• Nex-N2-Pro is post-trained on Qwen3.5-397B-A17B; Nex-N2-mini on Qwen3.5-35B-A3B.
• There is no Qwen3.6 at 397B; it only shipped as 35B-A3B and 27B. So "Nex-Pro + Qwen3.6" is undefined (you can't blend a 397B MoE with a 35B one).
• But Qwen3.6-35B-A3B keeps the exact Qwen3.5-35B skeleton (hidden 2048, 40 layers, 256 experts, vocab 248320, same modeling class). So Nex-N2-mini ⊕ Qwen3.6-35B-A3B is a real, defined merge: Nex blended with the newer base.

Auth (once)

pip install -r requirements.txt

# Modal token: https://modal.com/settings/tokens (or run: modal token new)
modal token set --token-id ak-XXXX --token-secret as-XXXX

# HF write token for your org, stored as a Modal secret so the cluster can push weights
modal secret create huggingface HF_TOKEN=hf_XXXX

Run it (Modal)

# free sanity check, no weights downloaded
python scripts/preflight.py --config configs/nex-mini-qwen36.yaml

# single 50/50 merge: download, merge, eval the merge + both parents (cached on a volume)
modal run modal_app.py::main --config configs/nex-mini-qwen36.yaml

# sweep blend ratios toward Nex and let the table pick the winner
modal run modal_app.py::sweep --weights-b 0.2,0.3,0.4

# pull results and print the table
modal volume get merge-theater results ./results
python scripts/compare.py results

# publish the winner (private by default; pass --no-private to go public)
modal run modal_app.py::publish_cmd --merged mini-b0.2 --repo-id PixelML/Mini-Merge-35B-A3B

Long jobs use modal run --detach so they finish server-side even if your laptop's connection drops. Note: a 35B in bf16 needs more than one 80GB GPU; the eval runs on a single H200 (141GB) at tensor_parallel_size=1 to avoid both the OOM and a flaky tensor-parallel comms timeout.

The 397B rematch (optional, not run here)

configs/rio-rematch-397b.yaml reproduces the original's ingredients (Nex-Pro + Qwen3.5-397B) with a smarter recipe (DARE-TIES). It's a ~$350–900 job and the 35B already makes the point, so it's left as an exercise.

The honest part (read before you claim SOTA)

1. A leaderboard bump is mostly a mirage. Merged models top boards through eval-format overfitting and benchmark contamination, not new capability.
2. A 35B does not out-think a 397B. The point isn't that this model is smarter than anything; it's that neither this nor the thing it mocks took the training that was claimed.
3. The flex is the cost: zero training, one config, a few dollars, fully reproducible, next to a "we trained an original 397B" claim.

What this is not

A model you trained. If you publish the weights, publish this repo with them: the config, the harness, the receipts. The guard that keeps it honest is preflight.py, which refuses to merge two models that don't share a skeleton. The merge itself is a plain weighted tensor average (scripts/merge_safetensors.py); there's no mergekit here, because its catalog didn't recognize this model's architecture. A linear merge really is a*W1 + b*W2.

Files

configs/nex-mini-qwen36.yaml   # Nex-mini + Qwen3.6 (the one that shipped)
configs/rio-rematch-397b.yaml  # Nex-Pro + Qwen3.5, DARE-TIES (the optional 397B version)
modal_app.py                   # download, merge, eval, publish, on Modal
scripts/merge_safetensors.py   # the actual merge: a weighted tensor average
scripts/preflight.py           # same-base guard (reads nested multimodal configs)
scripts/compare.py             # tabulate parents vs merged
scripts/evaluate.sh            # local lm-eval (hf or vllm)

---

Built end to end by Claude Code (Anthropic's agent, Opus) running autonomously overnight: the merge script, the Modal pipeline, eight dependency fixes, the ratio sweep, the published model, and the writeups. A human picked the targets and made the calls.