FINAL-Bench / Darwin-4B-Chimera

Darwin-4B-Chimera

A 4B Korean-reasoning model built with VIDRAFT's Chimera technology.

Most models get better by getting bigger. This one got better by combining — and then refining itself.

Why Chimera matters

Merging models usually means losing something. Average two networks and you get a compromise: each parent's edge is blunted, and the child is worse than both at what each did best. That is why "model merging" has a reputation as a cheap trick.

Chimera is built on the opposite premise.

1. The parents survive intact

Chimera fuses components from models of different families and different strengths while preserving what each parent knew. Knowledge is carried over, not averaged away. What you merge in is what you keep.

2. The result cannot be reduced to any single parent

A Chimera individual is not "model A with a bit of B." It sits at a point in model space that no parent reaches alone , and it cannot be recovered from any one of them. That non-reducibility is the entire point — it is what makes a lineage an asset rather than a re-skin.

3. No additional pretraining

A new generation is a fusion plus an evaluation , not a training run measured in GPU-months. Iteration collapses from months to days — which means the space of viable combinations actually gets explored instead of theorized about.

4. It compounds

Chimera individuals become parents. Strengths accumulate across generations: parent → child → grandchild. Capability is grown , not purchased.

5. It answers the real constraint

Frontier capability is gated by capital: tens of thousands of GPUs, months of training, hundreds of millions of dollars. Chimera attacks that gate directly — structural evolution instead of brute-force scale . It is how a small team competes on method rather than on budget.

What that buys you at 4B

This model is a late-generation Chimera individual, refined further by VIDRAFT's own reinforcement fine-tuning pipeline.

KMMLU · 6 subjects · 240 held-out items · greedy decoding

+5.4pp on Korean knowledge reasoning with zero parameter growth — 4.02B before, 4.02B after. The gain came from what the model learned from itself, not from more weights. Same architecture, same size, same inference cost: strictly better.

That matters because 4B is the size that actually ships. It runs on a single consumer GPU, on-premise, inside an air-gapped network — the places where frontier APIs cannot go.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

mid = "FINAL-Bench/Darwin-4B-Chimera"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype="auto", device_map="auto")

msgs = [{"role": "user", "content": "한국의 삼권분립을 간단히 설명해줘."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

What is open, and what is not

The weights are open and the numbers are stated with the exact evaluation setup , so anyone can reproduce the result above rather than take our word for it.

The internal design of Chimera fusion and of the refinement pipeline — component selection, data selection criteria, training configuration, routing — is VIDRAFT proprietary .

Results are open. The recipe is not.

Links

• Darwin family — https://huggingface.co/collections/FINAL-Bench/darwin-family
• Method paper (evolutionary merging) — arXiv 2605.14386
• VIDRAFT — https://vidraft.net

License

gemma — the Chimera lineage includes Gemma-derived components, so the Gemma license applies. Review the Gemma Terms of Use before use.