🔬 AI Research Digest
📅 Week of Sep 2–Sep 8, 2026
━━━━━━━━━━━━━━━━━━━━━━━━

1. 🌍 SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
Authors/Org: Junchao Huang, Guian Fang, Shengju Qian et al. | arXiv: 2609.02886
Bottleneck solved: No unified open framework existed for building interactive video world models that generalize across diverse environments and support long-horizon real-time rollouts.
SolarWM ships a fully open pipeline — 1.43M canonical clips (25+ TB) across real-world, synthetic, and game environments — plus a model family ranging from 5B to 33B parameters. For teams building simulation, robotics, or embodied AI products, this is now the go-to open foundation for world modeling at scale.
🔗 SolarWM on arXiv

━━━━━━━━━━━━━━━━━━━━━━━━

2. ⚡ FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution
Authors/Org: Shuo Yang, Xiaoze Fan et al. (UC Berkeley) | arXiv: 2608.16157
Bottleneck solved: Running frontier-scale Mixture-of-Experts models (290B+) locally on consumer hardware has been prohibitively slow or simply impossible due to fixed memory and bandwidth constraints.
FreeToken treats GPUs, CPUs, host RAM, and interconnects as a unified elastic platform — dynamically mapping computation to whatever hardware is available — achieving 39.3 tokens/s on a 35B model with just an 8GB laptop GPU. Developers can now run powerful open-weight models locally without datacenter-class hardware.
🔗 FreeToken on arXiv

━━━━━━━━━━━━━━━━━━━━━━━━

3. 🧪 nanochat — Full-Stack LLM Training Pipeline for $100
Authors/Org: Andrej Karpathy | GitHub: karpathy/nanochat
Bottleneck solved: Prior minimal repos like nanoGPT only covered pretraining, leaving a gap in accessible, end-to-end references for the full LLM development cycle.
nanochat packs tokenization, pretraining, finetuning, evaluation, inference, and a chat UI into ~8,000 lines of dependency-minimal code — trainable on 8 H100s in ~4 hours for under $100. An essential read for any ML engineer who wants to truly understand how ChatGPT-style models are built from scratch.
🔗 nanochat on GitHub

━━━━━━━━━━━━━━━━━━━━━━━━
💡 Stay curious. Read the papers.

For More: @kdnuggets @datasciencechats