Thox-ai/ThoxMini-3B-Space
0
⚠️ Org runtime paused. Temporary live demo: `tommytracx/ThoxMini-3B-Space`
ThoxMini-3B - free CPU tier
meta-llama/Llama-3.2-3B-Instruct + the THOX factuality LoRA trained in thoxllm-factory, on free cpu-basic.
The adapter is THOX-trained. The base model is openly licensed and named above; that derivation is not affiliation with or endorsement by its authors.
This adapter (thox-mini-3b-facts-hard-lora) is the only one in the fleet to PASS the product factuality gate outright - 3.50 against a 3.0 threshold.
Why not ThoxForge-7B
It was the first choice and is the better model. Without an active PRO subscription there is no ZeroGPU quota, so this Space runs on free cpu-basic (~16 GB RAM), and a 7B in bf16 is ~15 GB before overhead - it OOMs on load. 3B fits.
Known limitations
- CPU generation is single-digit tokens/sec. This tier is for a short scripted answer. The fast path is on-device, not here.
- The company-facts gate does not pass ungrounded. Ask about products, not about the cap table.
Cost
$0. Paid GPU is reserved for training runs.