Meta ships Muse Glimmer: a 30B Apache-2.0 agent that runs on one GPU
A capable local model, free for commercial use — plus OpenAI's vetted-only cyber model and where Qwen3.8's open weights actually landed.
Copy markdown
One 24GB GPU, no cloud, no tokens to buy
Muse Glimmer is a dense 30B model whose K-quant fits in 24GB of VRAM — a single RTX 4090/5090 — and generates 233 tok/s with DFlash speculative decoding (74.9 without). The day-one Hugging Face drop ships BF16 weights, GGUF k-quants, and ExecuTorch builds, so a genuinely capable agent runs locally at zero marginal cost.
Apache 2.0 — Meta drops the Llama license
Unlike recent Llama community terms, Muse Glimmer ships under a clean Apache 2.0 license: commercial use, fine-tunes, and redistribution with no strings attached. It's a 131K-context multimodal model (text and image in, text out) and marks Meta's first permissive open-weights release in a while.
Tuned for agents, not chat — build this weekend
It's benchmarked for the work agents actually do: MCP Atlas 75.5, SWE-Bench Pro 51.2 (leads its class), SWE-Bench Verified 77.2, AIME 2026 94.7 — though it trails on OSWorld (65.9) and TerminalBench (60.7). What you could build: a fully offline desktop agent that reads screenshots and makes schema-based tool calls, no API key required.
OpenAI's GPT-5.6-Cyber ships, but behind a gate
OpenAI released a cybersecurity-tuned model through its vetted-only Daybreak Red program; it has already surfaced previously unknown bugs in Chrome's V8 engine. Pricing runs $12.50/$75 per million tokens — roughly 2.5x standard Sol — and hardware security keys become mandatory Sept 1. Frontier defensive capability, but not open access.
Update: Qwen3.8-Max weights landed — the 27B didn't
The release builders were watching has a twist: the Qwen3.8-Max (2.4T) weights quietly went live on Hugging Face Aug 8, but under a bespoke qwen3.8-max license (not Apache) with unresolved geo-restrictions. The Qwen3.8-27B checkpoint most people actually want to run locally is still missing.