Multiplayer Labs
Research & technical reports
What we build, measured. Reports on model fusion, local inference, and verifier-driven autonomous agents — pick a paper to read the full write-up.
Mini Whale 1 12B
Fusing Qwen3-4B with 260 DeepSeek coding experts for consumer-GPU inference at 10 tok/s on a single RTX 3060 12GB.
A Mixture-of-Experts language model built on a fusion principle: a Qwen3-4B host augmented with SwiGLU coding experts extracted from DeepSeek-V4-Flash, low-rank bridge layers, a sqrt-softplus router, and residual-safe gating — fine-tuned with QLoRA and served with DSpark speculative decoding.
Sakura Island
Can a 1.5B local model improve itself without a human in the loop? Five self-improvement loops took it from 24% to 70.88% on HumanEval — with an autonomous Sakura Island build as the showcase.
A quantized 1.5B-parameter local model climbed from 24% to 70.88% on the 70 HumanEval tasks tested across five verified self-improvement loops, then authored a complete interactive voxel Sakura Island while a trusted harness handled parsing, compilation, browser execution, and scoring. No human intervention between submission and terminal verdict: 37.811 seconds, score 0.9439.
Local TTC
Building RSI and test-time compute with small models on local hardware — a 1.5B model learned to generate a 21-building NYC cityscape over 26 autonomous sessions, with no cloud and no API.
Over three days, a quantized 1.5B-parameter local model learned to generate a detailed three.js NYC cityscape through autonomous test-time compute. Token output grew from 124 to 358 per answer. LoRA training ran locally with 1.78M trainable parameters and 213K total tokens. The full pipeline — baseline, research, dataset, training, adapted artifact — completed end-to-end with zero human intervention.
New reports are published as experiments land. Every claim ships with its method, measurements, and limits.