Victoria: a 75% Terminal-Bench coding model in 49 GB

Farpoint's open model release: Qwen3.8-Flash-Next pruned 44% and retrained to recover the loss, scoring 75.3% on Terminal-Bench 2.1 in 49 GB of GPU memory, level with Qwen's dense 27B while reading 8 GB of weights per token against its 13.5 GB at the same 4-bit precision. Free for all users.