Howdy, CompactAI-O is launching a tiny Model Golf, and the winner walks away with $50 in RunPod credits. Monthly. Every month. Show up, build, somebody wins.
What it is
Build the best language model you can under 100 million parameters, with at least a 1028-token context window. That's it. Any architecture, any tokenizer, any training scheme you can dream up at 3am. The only catch is it's gotta be open source (MIT, GPL, Apache, AGPL) take your pick.
It scratches the same itch as a Kaggle comp without the dataset\leaderboard nonsense. No fixed benchmark to game. No llama.cpp compatibility hoops. If you wanna train a 50M-param MoE with five experts and a tokenizer built on cookbooks, you can do that. Nothing stopping you.
The rules are listed in the discord and on the organization page if you're interested.
Why $50????
It's symbolic. It ain't gonna make anyone rich. But it's enough to cover a weekend of GPU time, enough to keep enthusiasts coming back, and not so much that it pulls in people who are just there for the money. Enthusiasts build interesting things. Interesting things move the field forward. A little incentive. I'd do it for $50 lol.
🧬 Darwin Family: Zero Gradient Steps, GPQA Diamond 88.89%
How far can we push LLM reasoning *without* training?
Our team at VIDRAFT submitted this paper to Daily Papers yesterday, and it's currently #3. Huge thanks to everyone who upvoted — sharing the core ideas below.
Darwin Family is a training-free evolutionary merging framework. By recombining the weight spaces of existing LLM checkpoints — with zero gradient-based training — it reaches frontier-level reasoning.
- 🏆 Darwin-28B-Opus: GPQA Diamond 88.89% - 💸 Zero gradient steps — not a single B200 or H200 hour needed - 🧬 Consistent gains across 4B → 35B scale - 🔀 Cross-architecture breeding between Transformer and Mamba families - 🔁 Stable recursive multi-generation evolution
#Three Core Mechanisms
① 14-dim Adaptive Merge Genome — fine-grained recombination at both component level (Attention / FFN / MLP / LayerNorm / Embedding) and block level, expanding the prior evolutionary-merge search space.
② MRI-Trust Fusion — we diagnose each layer's reasoning contribution via an **MRI (Model Reasoning Importance)** signal and fuse it with evolutionary search through a **learnable trust parameter**. Trust the diagnostic too much and search collapses; ignore it and search becomes inefficient — Darwin learns the balance from data.
③ Architecture Mapper — weight-space breeding across heterogeneous families. Attention × SSM crossover actually works.
Why It Matters > Diagnose latent capabilities already encoded in open checkpoints, > and recombine them — no gradients required.