Gemma-4-26B-A4B优化模型大幅降低推理成本

BottleCapAI发布了ThinkingCap系列的第五个模型，基于Google的Gemma-4-26B-A4B稀疏专家混合模型进行优化1。该模型在保持精度基本不变的前提下，显著降低了推理所需的计算成本。

新模型在12项基准测试中的表现显示，思考tokens的使用量平均减少了38.9%，其中在LongBench v2上最高可减少69.1%1。精度方面，平均仅下降0.15个百分点，大多数结果均在噪声范围内，表明优化过程未对模型性能造成实质影响1。此外，模型的循环现象也得到改善，从原始模型的3.64%降至0.76%1。评估采用的基准包括AIME 2026、HMMT、GPQA-Diamond、MMLU-Pro、LiveCodeBench等1。

ThinkingCap PRO版本进一步优化了成本指标，使用的思考tokens平均减少62%，最高可达87.6%1。该模型基于26B参数架构，其中约4B为活跃参数的稀疏混合专家结构1。

BottleCapAI has unveiled the fifth model in its ThinkingCap series, a fine-tuned version of Google's Gemma-4-26B-A4B sparse mixture-of-experts model designed to optimize reasoning efficiency 1. The new model achieves substantial reductions in thinking tokens while maintaining comparable accuracy across multiple benchmarks 1.

The optimized model cuts thinking tokens by an average of 38.9%, with improvements reaching as high as 69.1% on LongBench v2, according to evaluations conducted at temperature 1.0 with top_p 0.95 and a generation limit of 65,536 tokens 1. Accuracy remained largely stable, declining by an average of only 0.15 percentage points across twelve benchmark tests—a figure within noise margins for most results 1. The model also reduced repetition in responses, lowering the occurrence rate from 3.64% in the original model to 0.76% 1.

BottleCapAI's ThinkingCap PRO variant achieved even more aggressive efficiency gains, using 62% fewer thinking tokens on average, with peak reductions of 87.6% 1. The underlying Gemma-4-26B-A4B model contains 26 billion parameters, though only approximately 4 billion are active due to its sparse expert architecture 1. Testing covered twelve evaluation benchmarks including AIME 2026, HMMT, GPQA-Diamond, MMLU-Pro, and LiveCodeBench 1.