Darwin: Reaching GPQA Diamond 90.9% by Evolutionary Merging, Not Pretraining
Merging instead of pretraining
The Darwin family is built by evolutionary merging of existing open models rather than from-scratch pretraining. The lineage draws on Gemma-4 and Qwen-3.5 open weights, and the released models are Apache-2.0. In practice that means the compute budget goes into search over merge recipes and parameter blends, not into a trillion-token pretraining run. For a team on roughly 24 GPUs, that is the difference between shipping and not shipping.
The GPQA Diamond result
On GPQA Diamond, Darwin reports 90.9%. In the snapshot taken in 2026-07, that placed it #3 globally and #1 in Korea on that specific benchmark, base-only. The reported ordering at the top:
| Rank | Model | GPQA Diamond |
|---|---|---|
| 1 | Kimi-K3 | 93.5 |
| 2 | GLM-5.2 | 91.2 |
| 3 | Darwin (VIDRAFT) | 90.9 |
Reportedly ranked below Darwin on that benchmark: DeepSeek-V4-Pro, Qwen3.5-397B, and Nvidia Nemotron-3-Ultra-550B. Only two Chinese frontier models, Kimi-K3 and GLM-5.2, lead. The notable part is the asymmetry in parameters and cost: a merged model derived from open weights landing next to models several times its size.
Why merging can work here
Evolutionary merging exploits a property of large pretrained models: much of their capability is already latent in the weights, and careful interpolation between complementary parents can preserve strengths while cancelling weaknesses. It is cheaper than pretraining, but it is bounded by the parents. A merge cannot invent a capability that none of its ancestors have. The honest way to read 90.9% is as strong recombination of Gemma-4 and Qwen-3.5 lineage capability, not as a novel training breakthrough.
Downloads and adoption
Darwin sits at the center of the roughly 1.18M cumulative downloads across the ecosystem on Hugging Face. That number needs context. A single community copy, ansulev/Darwin-9B-NEG, accounts for ~561K of it, while the original FINAL-Bench/Darwin-9B-NEG is near 2K; the copy out-downloaded the source by about 258x. Most of the rest sits on third-party mirror and quant accounts such as mradermacher and bartowski. This is real community adoption, and it is worth being explicit that a community copy, not a first-party repo, drove much of the total.
Honest scope
The 90.9% is single-benchmark and base-only. GPQA Diamond measures graduate-level multiple-choice reasoning; it does not settle instruction following, long-context behavior, safety, or agentic use. The global ranking is a 2026-07 snapshot on that one benchmark and will move. Darwin is a merge of open models, so its ceiling is tied to its Gemma-4 and Qwen-3.5 parents, under Apache-2.0 provenance. Download figures mix first-party repos, mirrors, quants, and one dominant copy. Treat Darwin as evidence that disciplined merging is competitive on a hard benchmark, not as a claim of frontier-scale training.
More: https://vidraft.net
Top comments (0)