Apex Intelligence has released its first model, Apex-35B-V1.0, alongside a demonstration in which that model helped train a smaller model to solve a puzzle. Apex says the smaller model's success rate rose from 11.7% before training to 24.4% on 512 unseen Sokoban levels. That comparison is with an untrained starting model, not another system that completed training.
How Apex-35B trained a puzzle solver
Sokoban is a grid puzzle in which a player pushes boxes onto target squares. In Apex's AI4AI-Bench task, Apex says Apex-35B-V1.0 trained Qwen2.5-3B-Instruct in two phases: four hours to explore training strategies, followed by 12 hours to retrain the Qwen model. Apex says the released model analyzed experiments and refined the training recipe. It reports that the retrained Qwen model solved 125 of 512 unseen levels, or 24.4%, compared with 11.7% for the untrained model.
Apex says a separate baseline agent exited after 101 minutes without submitting a model or completing a training step. Apex's announcement page does not link to an independent reproduction of the result.
Downloads and benchmark limits
The Hugging Face card lists the weights under the Apache-2.0 license and describes a sparse mixture-of-experts model with 35 billion parameters in total and about 3 billion active for each token. Its BF16 checkpoint is about 72 GB; the card also lists quantized GGUF versions. Apex says the model is based on Qwen3.6-35B-A3B and was further trained on proprietary data. The checkpoint is downloadable, but its full-precision file still requires substantial storage and suitable hardware.
Apex's other benchmark comparisons are also company-run and vary by task. The Sokoban result is a specific example of AI-assisted model training; by itself, it does not show that the approach can reliably improve models across tasks.
Sources
