A 35B Model That Runs on Your Phone Was Bred, Not Trained

A 28-year-old computer engineering student on leave from his university in Seoul just shipped a 35-billion-parameter model that runs on an iPhone. No cloud, no CUDA, no fork of anything. Kim Min-sik's startup, VIDRAFT, published POCKET-35B this week, and the pitch is blunt: stock llama.cpp, a laptop with no graphics card, and reasoning performance the company says holds up against models many times its size.

The twist is how VIDRAFT got there. POCKET wasn't trained. It was bred.

TL;DR

VIDRAFT, a Seoul AI Hub startup, released POCKET-35B, a 35B-parameter Mixture-of-Experts model that runs on CPUs, mini-PCs, and phones with no GPU
POCKET is a quantized descendant of Darwin-36B-Opus, which VIDRAFT built by gradient-free "breeding" two existing Qwen models rather than training from scratch
Darwin-36B-Opus ranks 5th of 1,252 models on Hugging Face's GPQA Diamond leaderboard at 88.4%, but that number comes from multi-pass stochastic voting - a single greedy pass scores 73.2%, identical to its unmerged base model
POCKET's compressed variants score lower still on the same benchmark, 60.1% to 68.7% depending on quantization, and get slower rather than faster during GPU prompt processing
The merging method, called Darwin V7, is documented in an arXiv paper and released under Apache 2.0

The Claims, Side by Side

Claim	VIDRAFT's number	The catch
GPQA Diamond rank, Darwin-36B-Opus	5th of 1,252 models, 88.4%	Requires multi-pass stochastic retry with majority voting
GPQA Diamond, single greedy pass	73.2%	Matches base Qwen3.6-35B-A3B exactly
POCKET-35B Q4_K_M, GPQA greedy	68.7%	Best-quality quant, still below the flagship
POCKET-35B Q2_K, GPQA greedy	60.1%	VIDRAFT's own "best value" recommendation
CPU generation speed vs Bonsai-27B	2.69x faster (27.0 vs 10.1 tok/s)	Measured on a 16-thread Xeon
H100 prompt processing vs dense rival	0.41x (slower)	MoE prefill wakes every expert, sparsity doesn't help

A Model Bred, Not Trained

VIDRAFT's core technology isn't POCKET, it's the engine that produced its parent. Called Darwin V7, the system treats existing open model checkpoints as a genetic pool and recombines their weight tensors without a single gradient step. Per-tensor compatibility analysis decides which components can merge safely, an automated recombination pass builds the offspring, and a multi-phase benchmark run verifies the result. VIDRAFT says the whole breeding cycle takes under 10 minutes on one GPU, a fraction of what pretraining a model this size normally costs.

Darwin-36B-Opus, the parent behind POCKET, came from crossing Qwen3.6-35B-A3B, Alibaba's hybrid Gated DeltaNet/attention MoE base, with a second model fine-tuned on Claude Opus 4.6 reasoning traces. The recombination logic biases toward the reasoning parent for components tied to chain-of-thought and toward the base parent everywhere else. The method, detailed in an arXiv paper titled "Darwin Family: MRI-Trust-Weighted Evolutionary Merging," calls this a 14-dimensional "adaptive merge genome," and it's now sitting in front of NeurIPS reviewers.

The Leaderboard Climb

The technique's real proof point predates POCKET by three months. On Hugging Face's GPQA Diamond leaderboard, a 1,252-model ranking of PhD-level science reasoning, VIDRAFT's Darwin family occupies every one of South Korea's five entries in the top 21, with Darwin-28B-Opus taking 3rd place at 88.89%, ahead of a dense 397-billion-parameter competitor roughly ten times its size. China holds the other 16 spots in that top tier.

Seoul skyline at night, the city where VIDRAFT is based VIDRAFT operates out of Seoul AI Hub, a city-backed startup facility, and became the only South Korean startup to get a direct collaboration proposal from Hugging Face in March 2026. Source: unsplash.com

"Given the cost and resources required to build large models from scratch, we needed a different approach. Combining models turned out to deliver real performance gains," VIDRAFT CEO Kim Min-sik said of the strategy.

VIDRAFT was founded in 2024, is backed by South Korea's National IT Industry Promotion Agency, and has spun its Darwin lineage into roughly 380 derivative models on the open-source leaderboard ecosystem. For a team this size to out-rank labs with dedicated pretraining budgets, on a scientific reasoning benchmark no less, is the kind of underdog result that gets attention. Whether it holds up depends entirely on how you read the number.

What They Measured, and What They Didn't

The Multi-Pass Gap

The 88.4% headline score isn't what Darwin-36B-Opus gets on the first try. VIDRAFT's own model card shows Pass 1, greedy decoding, lands at 73.2%. Pass 2, which retries with temperature-based sampling and takes the majority vote across attempts, recovers another 15.2 points. That's a legitimate technique, plenty of labs report best-of-N numbers, but it means the leaderboard-topping figure describes a different, more expensive inference process than a single query. And the greedy number, the one closest to how most people actually run a chatbot, is identical to the unmerged Qwen3.6-35B-A3B base model. The breeding process didn't move that needle at all.

The Compression Cost

POCKET inherits the same ceiling and loses more on the way down. Its Q4_K_M build, the largest at 21.2GB, scores 68.7% greedy. The 12.9GB Q2_K variant VIDRAFT recommends as the "best value" option drops to 60.1%. When a Hugging Face commenter, Dipankar Sarkar, asked why POCKET loses to a dense competitor on H100 prompt processing (0.41x) but wins on the same operation on a MacBook M3 Pro, the mismatch points to something POCKET's own documentation admits: MoE sparsity helps token generation, where only 8 of 256 experts activate, but prompt processing wakes every expert regardless, erasing the advantage on hardware built for dense matrix throughput.

None of this means the numbers are fabricated. VIDRAFT publishes its quantization variants, its perplexity scores per language, and its benchmark harness openly, and the whole stack runs on stock llama.cpp anyone can audit. It means the marketing framing (top-5 on a leaderboard of 1,252 models) and the practical framing (what a single query on your phone actually returns) are two different claims wearing the same benchmark number.

Abstract rendering of a DNA double helix, evoking the "breeding" metaphor VIDRAFT uses for its merging technique VIDRAFT frames Darwin's weight-recombination process as evolutionary "breeding" rather than training - a framing this piece borrows because it's technically apt, not because it's marketing copy. Source: unsplash.com

Should You Care?

If the goal is a truly capable reasoning model that fits on a phone or a GPU-less mini PC, POCKET-35B is a real option, not vaporware. Apache 2.0 licensing, honest per-quant benchmarks, and 2-3x CPU throughput over a comparably sized Bonsai-27B are useful wins for anyone doing on-device inference. Just don't treat the GPQA Diamond rank as a description of what you'll get from a single prompt on a 16GB laptop. Read it the way you'd read any contested leaderboard result: as the best case under generous sampling, not the expected one.

For VIDRAFT, the more durable story isn't the leaderboard slot, it's the method. Breeding a 35B model from existing checkpoints in under 10 minutes on a single GPU, then compressing the result for phones, is a genuinely different way to compete with labs spending hundreds of millions on pretraining. Whether Darwin V7 keeps producing gains as the genetic pool of available open models gets picked over is the open question worth watching, not this month's leaderboard screenshot.

Sources: