You can deploy this model and start earning money today!

FINAL-Bench/
Darwin-218B-Expert

Name: Darwin-218B-Expert
Availability: InStock

Views

No views yet

🌌 Darwin-218B-Expert

VIDRAFT FINAL-Bench Darwin family multi-domain Expert flagship — 218B parameter Cohere2 MoE base with chemistry + biology domain knowledge integrated via task arithmetic on Korean-aligned base.

A multi-domain "Expert" derivative of the Darwin-218B family. Built by composing chemistry and biology domain deltas onto the Korean-aligned base, this v1 release preserves Korean fluency while adding scientific reasoning in chemistry and biology.

Lineage

CohereLabs/command-a-plus-05-2026-bf16     (base 218B MoE, ~25B active)
              ↓ Korean LoRA merge
       Darwin-218B-kr                       (Korean-aligned base)
              ↓ Chem LoRA merge        ↓ Bio LoRA merge
   Darwin-218B-Expert-Chem      Darwin-218B-Expert-Bio
                       ↓ task arithmetic ↓
              Darwin-218B-Expert  ← THIS MODEL (v1 = Chem + Bio)

Construction (v1): task arithmetic on the Korean base

W_Expert = W_Chem + W_Bio − W_kr
        = W_kr + ΔChem + ΔBio

Both chemistry and biology deltas preserved at 100% (no dilution), with Korean fluency inherited from the kr base.

Domain Capabilities

Domain	Coverage	Source
Korean (한국어)	Native fluency	Inherited from `Darwin-218B-kr` base
Chemistry	Organic, spectroscopy, physical, inorganic, analytical, special (6-domain SFT)	`Darwin-218B-Expert-Chem` delta (Opus-distilled, anti-contamination)
Biology	Cellular, molecular, biochem, ecology, evolution (6-domain SFT)	`Darwin-218B-Expert-Bio` delta (Opus-distilled, anti-contamination)
General	All capabilities from Cohere Command A+ retained	Underlying base

Architecture

Item	Value
Parameters	218B total / ~25B active (MoE)
Architecture	`Cohere2VisionForConditionalGeneration` (multimodal-capable, text-primary)
Experts	128 (Cohere2 MoE structure)
Precision	BF16
Tokenizer	Cohere2 (vocab 256K)
Languages	English, Korean
Context	65,536 tokens

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "FINAL-Bench/Darwin-218B-Expert",
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-218B-Expert")

messages = [
    {"role": "user", "content": "Explain the mechanism of SN2 reaction step by step. Then describe how mRNA splicing maintains reading frame fidelity."}
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Serving (vLLM, recommended):

bash

vllm serve FINAL-Bench/Darwin-218B-Expert \
    --tensor-parallel-size 8 \
    --dtype bfloat16 \
    --max-model-len 65536 \
    --trust-remote-code

Requires vLLM ≥ 0.21.0 with Cohere2VisionForConditionalGeneration support.

Methodology — Task Arithmetic

Unlike sibling-model averaging (which dilutes expertise), this model uses task vectors (Ilharco et al., 2022) to compose domain knowledge:

Train domain-specific LoRA on a shared base (Darwin-218B-kr)
Each merged expert (Expert-Chem, Expert-Bio) encodes a domain delta from the base
Add deltas to the base: W_final = W_base + Σ Δ_domains

This is mathematically equivalent to applying multiple LoRA adapters simultaneously, but produces a single dense checkpoint requiring no runtime adapter loading.

Properties:

Both expert deltas preserved at 100% magnitude
No interference dilution (unlike weighted blends)
Korean base layer intact

Version Roadmap

Version	Domains	Status
v1	Korean + Chemistry + Biology	✅ Current
v2	+ Mathematics	⬜ Planned
v3	+ Code	⬜ Planned
v4	+ Physics, Law	⬜ Planned

Each domain delta is added via task arithmetic on the shared Darwin-218B-kr base.

License

Apache License 2.0

Built upon CohereLabs/command-a-plus-05-2026-bf16 (Apache-2.0) and Darwin-218B-kr (Apache-2.0). Both upstream components are permissively licensed; users are responsible for compliance with upstream terms.

Contributors

Lead Architect & Developer 장재원 (Jaewon Jang) — CTO, VIDRAFT Multi-domain Expert composition design, task-arithmetic merge pipeline, domain SFT distillation pipelines.

Organization VIDRAFT / FINAL-Bench https://huggingface.co/FINAL-Bench

Citation

bibtex

@misc{darwin-218b-expert-2026,
  title  = {Darwin-218B-Expert: Multi-Domain Expert via Task Arithmetic on Korean-Aligned 218B MoE},
  author = {Jang, Jaewon and {VIDRAFT FINAL-Bench Team}},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-218B-Expert}}
}

References

Task Arithmetic: Ilharco et al., "Editing Models with Task Arithmetic", ICLR 2023 — arXiv:2212.04089
TIES-Merging: Yadav et al., "TIES-Merging: Resolving Interference When Merging Models", NeurIPS 2023 — arXiv:2306.01708
Cohere Command A+ (base): CohereLabs/command-a-plus-05-2026-bf16
Darwin-218B-kr (Korean base): private — FINAL-Bench/Darwin-218B-kr
Expert sources: private — FINAL-Bench/Darwin-218B-Expert-Chem, FINAL-Bench/Darwin-218B-Expert-Bio