Ling 3.0 Tiny

6gb here. Put Ling tiny in a harness and go wild

Why I find it surprising is I can run Gemma 4 12b and Gemma 4 26b both QAT versions with some offloading at the 20-25 tok/sec range. Ling 3 Tiny should surely be much faster than them.

Ling-3.0-Tiny gives me only 30 t/s while Ling-mini-2.0 gives me 50-60 t/s on CPU-only inference itself. Same with GPU-CUDA. 70-80 t/s vs 150+ t/s

LFM2.5-2.6B - It turned out significantly better for me. I tested it a bit, though, specifically in terms of extracting and finding the desired text note in obsidian and ling 3.0 tiny shows more tool calls issues

Ling-tiny is really cool

I have had good luck using ling-3.0-tiny for the OpenZim MCP server for example, that model will work its way through the damn encyclopedia and pull out the answers (and I can verify where they came from). Sure Qwen3.8 does it just as well,

I find Ling 3.0 tiny particularly interesting as it looks really nice for a tiny model with 7.9B total parameters, with only 1.3B parameters activated per token

smarter than many 30b parameter models