0.8B model with 9 LoRA adapters gives 38x faster decisions

A developer released Jeff-Qwen3.5-0.8B plus 9 task-specific LoRA adapters that answer common agent decisions first, passing only uncertain queries to Qwen3.8-27B. On an M4 Max, accuracy rose from 86.6% to 95.3% and time per decision fell from 8.1s to 0.25s, for under 2 GB of extra memory.

• Five Qwen3.8-27B models beat Fable 5 on LiveCodeBench Hard2026-11-01
• Nanbeige releases Nanbeige4.3-33B model2026-10-25
• Five Qwen3.8-27B models match Fable 5 in benchmark tests2026-10-25
• ThinkingBox uses database checks as RL rewards2026-10-22
• Developer builds Pokemon 2v2 multi-agent LLM test environment2026-10-22