The Hub is quiet today. Two items stand out. One is a new streaming pack for a 35B MoE that runs on a Mac with about 3 GiB of active RAM. The other is a community blog that maps how labs are building sandboxes for agentic RL.

On the Hub

Edge0-35B-A3B-preview: Qwen3.5-MoE 35B, 3 GiB active, Apple Silicon

Edge0-AI put Edge0-35B-A3B-preview on the Hub on September 8 and updated it on September 10. The pack is Apache-2.0 and built for MLX on Apple Silicon. It streams Qwen/Qwen3.5-MoE-35B-A3B experts from SSD with int4 weights, a LoRA adapter and a prerouter. 1

The card lists 40 layers, 256 experts with 4 active, and hidden size 2048. The Hub tree is about 19.74 GB. On a Mac mini M4 Pro active memory is about 2.9 GiB. 2

Decode runs 14.9 to 17.7 tok/s. Prefill runs 113 to 140 tok/s cold to warm. The prerouter is claimed to add up to 59 percent decode speed. 2

Card-claimed OpenCompass numbers show a small loss versus fp16 base. AIME 2026 is 86.6 versus 92.7. HumanEval is 90.9 versus 95.1. The average is 79.2 versus 83.2. 3

The framework is MLX 0.30.6 and mlx-lm 0.31.0, Python 3.10 plus, macOS Apple Silicon only. CUDA is on the roadmap. A sibling Edge0-8B-A1B-preview based on inclusionAI/Ling-3.0-tiny-base is also present. 3

Sergio Paniego: one sandbox per agentic-RL rollout (Sep 11 HF blog)

Hugging Face community author Sergio Paniego published One sandbox per rollout, or how labs run RL for agents in 2026 on September 11. It is class 4 of the Training Agents series. The post surveys fifteen reports from thirteen labs between October 2025 and September 2026. 4

Labs named include GLM-5 and 5.3, Kimi K3, MiniMax, Nemotron 3 Ultra, OLMo 3, DeepSeek, Qwen3.8, Inkling, LFM2.5, MAI-Thinking-1, GPT-5.6 and GPT-6 Astra, Cursor Composer and Meta Code World Model. The pattern is one dedicated sandbox per rollout attempt. 4

Cursor runs hundreds of thousands of concurrent sandboxed coding environments. Kimi K3 uses resumable microVM sandboxes for million token agentic RL. 4

GLM-5.3 notes that as agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment. The post contrasts white-box harnesses that rebuild the environment with black-box proxies of shipped agents. Public stacks mentioned include OpenEnv, TRL, Harbor, verifiers, slime, verl and HF Sandboxes experimental. 4