Qwenium
A self-contained C++ inference engine for the Qwen and Gemma model families on ggml, built around auditable answers that cite which parts of the input they used and which they never read.
Qwenium serves up to 10 concurrent users from a single 32 GB Mac, supporting vision input, speculative decoding with a trained MTP head, grammar-constrained output, portable session snapshots and warm conversational serving, with custom fused DeltaNet kernels running through Metal.
Every answer comes with per-value source citations for what it read, an explicit statement of what it never consulted, and a byte-identical re-run for verification, which the project describes as a calibrated capability rather than a blanket guarantee for every model it hosts.
With flash attention and a persistent graph enabled, its decode speed is within 21 percent of llama.cpp on the same ggml revision, and its forward pass matches HuggingFace transformers token-for-token.