Bonsai-8B vs OpenMythos

PrismML, a Caltech spinout, has shipped Bonsai-8B — the first 1-bit large language model that claims genuine benchmark parity with leading full-precision 8B instruct models while fitting entirely in 1.15 GB of RAM. It runs natively on Apple Silicon via MLX and on NVIDIA GPUs via llama.cpp without any quantization post-processing. The breakthrough here isn't just size — it's efficiency. PrismML reports approximately 4-5x better energy efficiency versus traditional 8B models, which matters enormously for mobile deployment, embedded systems, and cost-sensitive inference at scale. The Apache 2.0 license means no commercial restrictions, and the team has published the full training methodology alongside the weights. Previous 1-bit LLM efforts (BitNet, etc.) delivered underwhelming benchmark performance at practical scales. Bonsai-8B claims that gap has finally closed. If the benchmarks replicate independently, this could be the model that makes "AI on every device" a 2026 reality rather than a 2028 roadmap item.

OpenMythos is a community-driven theoretical reconstruction of Claude Mythos's suspected architecture, implementing a Recurrent-Depth Transformer (RDT) — a looped transformer that recycles layers multiple times per forward pass for deeper reasoning without massive parameter growth. The project drew 10,100 GitHub stars in its first week, reflecting intense developer curiosity about what's powering Anthropic's latest generation models. The architecture has three stages: a Prelude (initial layers), a Recurrent Block (looped up to 32 times with shared weights), and a Coda (final layers). Rather than stacking hundreds of unique layers, the recurrent block runs the same weights multiple times with learned injection parameters updating hidden states between loops — enabling implicit chain-of-thought reasoning in continuous latent space without generating intermediate tokens. The project supports Grouped Query Attention (GQA) with optional Flash Attention 2, Multi-Latent Attention (MLA), and sparse MoE with routed and shared experts. Model scales range from 1B to 1T parameters. The key claim is that RDT achieves reasoning depth comparable to fixed-depth models with far more parameters, since computational complexity scales with loop iterations rather than layer count. This would explain how Claude Mythos achieves strong reasoning performance without the extreme parameter counts of brute-force scaling — though Anthropic has neither confirmed nor denied the architecture.