DeepSeek-V4-Flash-0731 'Low' Reasoning Effort Mode Unexpectedly Uses More Tokens Than 'High'
Empirical testing of the DeepSeek-V4-Flash-0731 model has revealed that its "Low" reasoning effort mode unexpectedly generates more reasoning and final tokens than the "High" effort mode. This counterintuitive behavior was observed both in local tests using Unsloth's UD-Q2_K_XL quantization and via DeepSeek's official API. This finding is crucial for developers looking to optimize API costs and latency, as selecting "Low" effort might unintentionally lead to higher token usage and costs compared to "High" effort. It also highlights potential inconsistencies in how reasoning effort parameters are implemented or interpreted by the model. In API tests, the "Low" mode averaged 1,349.2 total tokens (889.6 reasoning) compared to only 481.5 total tokens (253.9 reasoning) for the "High" mode. Additionally, the tester noted a significant bug on OpenRouter that currently breaks the reasoning effort modes for this model.
## BACKGROUND
Reasoning models, or Large Reasoning Models (LRMs), are trained to perform multi-step logical thinking before generating a final answer. Modern LLM APIs allow users to adjust "reasoning effort" parameters to control how much computational budget or "thinking time" the model should allocate to a task. Unsloth's Dynamic v2.0 GGUF (such as UD-Q2_K_XL) is an advanced quantization method designed to run LLMs locally with high accuracy and reduced memory footprint.