Qwen 7B and Mistral Nemo 12B passed the classroom speed gate in round 2 and were retained as candidates pending faculty review¶

Claim · All claims
Evidence · 1 study (2 entries) · design r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Qwen 7B and Mistral Nemo 12B both passed 12/12 probes with latency ratios of 1.01 and 1.24, and Qwen completeness was 0.93 versus 0.91 for 8B. → Sidney Shapiro and Joshua Lindemann 2026 q2 i? Round 2 counters categorical rejection of larger models: Nemo 12B completed at 1.24 times its fresh baseline. → Sidney Shapiro and Joshua Lindemann 2026 (2)

Evidence¶

Sidney Shapiro and Joshua Lindemann 2026¶

Sidney Shapiro and Joshua Lindemann. (2026). On-Premises Multi-Course RAG Tutoring for Business Education: Hardware–Software Trade-offs in a Campus AI Tutor. arXiv. https://arxiv.org/abs/2610.02510 doi:10.48550/arxiv.2610.02510

q2 · i? · design · r3

Bake-off round 2 on Node A with non-thinking candidates and a fresh 8B baseline; the decision summary reports "latency ratios were 1.01 and 1.24" and a small completeness-proxy difference, and both were retained as candidates for paired review.

"Both passed 12/12 probes; latency ratios were 1.01 and 1.24. Qwen completeness was 0.93 versus 0.91 for 8B."

Sidney Shapiro and Joshua Lindemann 2026 (2)¶

Sidney Shapiro and Joshua Lindemann. (2026). On-Premises Multi-Course RAG Tutoring for Business Education: Hardware–Software Trade-offs in a Campus AI Tutor. arXiv. https://arxiv.org/abs/2610.02510

q2 · i? · design · r3

The results section states the round-2 outcome in the authors' words: "Nemo 12B completed at 1.24 times its fresh baseline", showing a larger model can fit the latency allowance.

"Round 2 provides a counterexample to rejecting larger models categorically: Nemo 12B completed at 1.24 times its fresh baseline."

Discussion¶

Related Claims¶

• Larger dense generators passed smoke probes but all exceeded the +30% classroom latency limit in bake-off round 1 — related
• Head-to-head comparison exposes a security-usability-latency tension: NeMo reaches 0% bypass at 16.22% FPR and ~1.5 s latency while Prompt Guard yields 38.48% bypass at 3.60% FPR — related
• The most energy-efficient tested configuration consumes 1.8 mWh per query, a 43.8% reduction relative to the Qwen-7B FP16 baseline — related
• NF4 4-bit quantization reduces VRAM usage by 38.7% for Qwen-7B and 37.9% for DeepSeek-MoE while maintaining practical response times — related