Glyd v0.30.0: download and run

Glyd v0.30.0 is out.

Download and run. glyd run MODEL:swift (or :kestrel) now downloads a ready-made checkpoint from Hugging Face and starts from it, instead of fitting the weights on your computer first. Ready now: Qwen3.8-27B, Qwen3.6-27B and Qwen3.5-27B, 9B, 4B, 2B and 0.8B, each at swift and kestrel, each with a card of what was measured.

curl -LsSf https://getglyd.com/install.sh | sh
glyd run Qwen/Qwen3.8-27B:swift

Nemotron 3 Super on one GPU. NVIDIA’s Nemotron 3 Super (120B, 12B active) runs at :hummingbird, about 3 bits per weight for its experts, from glyd/nemotron-3-hummingbird (48.7 GB). When the GPU cannot hold the whole model, Glyd keeps the most-used parts on the GPU and the rest in the computer’s memory, and the CPU does the work for the parts that are not on the GPU; with less memory than the model, the rest is read from the SSD as needed.

glyd run nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16:hummingbird

One user, measured:

Machine	Tokens a second
RTX 4090	43.6
RTX 3090	37.0
An RTX 4090 held to a 16 GB card’s memory	35.4
32 GB of RAM, a fast NVMe SSD, RTX 4090	17.7

GSM8K 97.0% (300 questions), MMLU-Pro 76.6% (222), not thinking.

Also: glyd serve returns tool calls from Qwen3.5, Qwen3.6, Qwen3.8 and Nemotron 3 as tool_calls, streamed or not; sampling picks the next token on the GPU; the levels are penguin, kestrel, swift and hummingbird.

Full details are in the release notes and the changelog; what is not done yet is on the known gaps.