Raw1 AI – 100% Local AI.

Raw1 AI Model Release News – September 30, 2026

AI MODEL MONITOR

AI Model Release Monitor — Daily Digest (2026-09-30)

2026-09-30

2 new model release(s) found.

CHINA / LOCAL MODELS

Hugging Face
A 27B Qwen3.8-based long-horizon self-improving agent that iteratively proposes, tests, reflects on, and revises solutions. It targets coding and machine-learning engineering as well as deep research, with a 262K-token context window.
Source: https://huggingface.co/BAAI/AREX-2
Hardware: ~60 GB VRAM at FP16/BF16, ~30 GB at INT8, or ~15 GB at 4-bit (GGUF/AWQ) including overhead; Q4 fits a high-end consumer GPU with 16–24 GB VRAM, while full precision requires a larger GPU/server.

OTHER RELEASES

Hugging Face Blog
An open foundation model for tabular classification and regression that predicts labels or numeric targets in one forward pass without per-dataset training or feature engineering. The three-size family tops out at 215M parameters and is available on Hugging Face and GitHub.
Source: https://huggingface.co/nvidia/Kumo-Tabular
Hardware: Lightweight: largest 215M variant is about 0.5 GB at FP16/BF16, 0.24 GB at INT8, or 0.12 GB at 4-bit including ~10% weight overhead; CPU/laptop RAM or a modest GPU is ample. HF API has no safetensors total, so sizing uses the 215M maximum reported by NVIDIA.
Full browsable, searchable model page
Source: https://drive.google.com/file/d/1NpL8_FQv2YM5fM2x8N9A5w8jgfi3xrjQ/view?usp=sharing
RAW1 AI — You don't need more employees to do repetitive work. You may need better automation.