The reasoning built into the models matter so much too. I recently swapped my Qwen3.6 27B to ThinkingLabs’ fine tune and it does what it publishes. I cut my token usage in half, which is a big deal since I only get ~20 TPS for token generation.