Darwin-36B-Opus: How VIDRAFT's Evolutionary Engine Bred a 36B Model That Rivals 397B Giants
TL;DR: Darwin-36B-Opus is a 36-billion-parameter Mixture-of-Experts model created not through conventional training, but through weight recombination of two parent models using VIDRAFT's Darwin V7 evolutionary engine. It scores 88.4% on GPQA Diamond — matching a 397B-parameter model — while activating only 3B parameters per forward pass. For engineers looking to deploy frontier-grade reasoning at a fraction of the compute cost, this is worth your attention.
What it is
Darwin-36B-Opus is a sparse Mixture-of-Experts (MoE) language model produced by FINAL-Bench / VIDRAFT_LAB. Key specs:
- Total parameters: 36B
- Active parameters per token: 3B (8 experts active out of 256 routed experts)
- Context window: 262,144 tokens
- Precision: BF16
- License: Apache 2.0 (commercially usable)
- Supported languages: English, Chinese, Korean, Japanese, German, French, Spanish, Russian, Arabic
The model is not a fine-tune in the traditional sense. It is the offspring of two parent models — both rooted in the Qwen3 MoE architecture — combined through principled weight recombination rather than gradient descent on new data.
How it works
Traditional model development involves weeks of GPU-cluster training, terabytes of data, and significant capital. Darwin V7, VIDRAFT's evolutionary engine, takes a different approach: model breeding.
The high-level process looks like this:
-
Select two parent models that share a compatible architecture. For Darwin-36B-Opus, the parents are:
- A base Qwen3-35B MoE model with hybrid attention and 256 routed experts
- A reasoning-distilled variant of the same architecture, fine-tuned on Claude Opus 4.6 chain-of-thought traces
Recombine weights automatically. Darwin V7 identifies which weight regions from each parent to inherit, aiming to preserve the reasoning behaviors of the distilled "mother" model while retaining the structural expert-topology precision of the "father" model.
Produce an offspring model that inherits complementary strengths from both parents — specifically, long chain-of-thought reasoning from one lineage and expert routing fidelity from the other.
The entire process reportedly completes in under one hour on a single GPU — a dramatic contrast to multi-week training runs.
Think of it as analogous to genetic crossover in biological evolution, applied to the weight tensors of neural networks. The key insight is that high-quality pretrained and fine-tuned models already encode enormous amounts of structured knowledge; evolutionary recombination is a way to explore the space of possible combinations without paying the full training cost each time.
Benchmarks & results
The headline number is 88.4% on GPQA Diamond — a rigorous 198-question benchmark covering graduate-level physics, chemistry, and biology problems. This is widely considered one of the harder public evals for reasoning models.
How that stacks up against publicly reported scores:
| Model | GPQA Diamond |
|---|---|
| TNSA/NGen-4-Pro | 91.1% |
| TNSA/NGen-4 | 90.1% |
| Darwin-36B-Opus (36B) | 88.4% |
| Qwen3.5-397B-A17B | ~88.4% (tie) |
| Kimi-K2.5 | 87.6% |
| Darwin-27B-Opus | 86.9% |
| Qwen3.5-122B-A10B | 86.6% |
| GLM-5.1 (744B) | below 86.6% |
Per the GPQA Diamond leaderboard dated August 23, 2026, Darwin-36B-Opus holds 3rd place overall — achieved with a model that is roughly 11× smaller by parameter count than the 397B model it ties.
How to try it
The source article links directly to Darwin-36B-Opus on Hugging Face, indicating the model is publicly available. You can access it here:
Hugging Face: Darwin-36B-Opus model page (follow the link in the source article)
To download via the Hugging Face CLI:
pip install huggingface_hub
huggingface-cli download <model-id-from-hf-page>
Replace
<model-id-from-hf-page>with the exact repository identifier shown on the Hugging Face listing. The source article does not print it inline, so verify it directly on the model card.
Given the Apache 2.0 license, you can use this model for commercial applications without royalty restrictions. With only 3B active parameters per forward pass, it is also deployable on hardware that would be impractical for a 397B dense model.
FAQ
Q: Is this actually training-free, or is there still some gradient computation involved?
A: Based on the source description, Darwin V7 performs weight recombination — not backpropagation on new data. The "under one hour on a single GPU" figure is consistent with inference-level or merging-level computation, not a training run. That said, the distilled parent model was itself trained via knowledge distillation, so prior training is embedded in the lineage.
Q: Does the evolutionary approach generalize, or is it specific to Qwen3-family architectures?
A: The source describes Darwin V7 as requiring architecturally compatible parent models. Both parents here share the same base architecture, which is why weight recombination is tractable. Whether Darwin V7 supports cross-family merges is not detailed in the public coverage.
Q: Can I use Darwin-36B-Opus for commercial products?
A: Yes. The model is released under Apache 2.0, which permits commercial use, modification, and redistribution with attribution.
Originally reported by vc.ru (러시아) (2026-08-23) — source article.
Top comments (0)