SAVRN's Take
Five models on our hub carry osl-3.0, and all five come from one publisher, Convergent Intelligence. We have not summarized what this license permits for commercial use; the license text governs. For an operator that puts one task ahead of every other: before the weights go into a rack that serves paying customers, someone on your side reads the full license and decides whether serving, modifying and redistributing the model fits the terms as written. Do that review before the purchase order, not after.
All five are text generation models on the Gemma3ForCausalLM architecture, and each is a weight-level merge of Gemma-3 (1B) and Qwen-3 (0.6B), tuned afterward for stepwise reasoning. The three we size, Qemma-Q14B, Qemma-Q1.7B and Qemma-GEI, hold about 1 billion parameters and need 2.4 GB at 16-bit, 1.2 GB at 8-bit and 0.6 GB at 4-bit, with the cheapest host we track at a single MI300X for $1.85 an hour. At that footprint, memory is not what drives the hardware choice. Context length is where they separate: Qemma-Q14B takes 524,288 tokens, Qemma-Q1.7B 262,144, Qemma-GEI 131,072, and Qemma-sft and Qemma-redux 32,768 each.
Downloads sit in a narrow band. Qemma-sft leads at 3,832 a month, Qemma-redux follows at 3,808, then Qemma-Q14B at 3,738, Qemma-Q1.7B at 3,691 and Qemma-GEI at 3,634. Read the license text first, then choose by context length; memory and host price are the same across the three we size.
SAVRN Research, 2026-09-18
