Qwen3.6-35B-A3B posts profit on FoodTruck Bench

FoodTruck Bench is a better sanity check for agentic models than most static coding evals, and Qwen3.6-35B-A3B clearing it with profit suggests the model’s planning loop is legitimately useful, not just benchmark-friendly.

• –Finishing 30 simulated days matters more than a flashy one-shot score because the benchmark punishes bad inventory, staffing, and cash-flow decisions
• –Ranking 11th while staying in the black puts it in the “actually operational” tier, not just “can complete tasks eventually” territory
• –The result is especially notable for a 35B-class MoE model, since it competes on endurance and economics against much larger systems
• –For developers building agents, this is a stronger signal than plain tool-use demos because it tests memory, persistence, and compounding decisions over time
• –The benchmark’s business-sim framing makes profitability the real criterion, which is closer to deployed agent work than standard QA-style evals