A small AI developer has released a version of Google's gemma-4-26B-A4B-it that reasons for far fewer tokens before answering, and the company says the trade-off in accuracy is minimal. The model, called ThinkingCap, cut thinking tokens by 38.9% on average across twelve benchmarks, with an average accuracy change of negative 0.15 percentage points. The developer, which publishes under the name bottlecapai, says it is the fifth model in its ThinkingCap series and the first built on a Google model.
The base model is a sparse mixture of experts with 26 billion parameters, of which about 4 billion are active for each token. According to the company, the training goal was narrow: get to the answer using fewer tokens and change nothing else. The developer says it set out to preserve knowledge, reasoning ability, answer quality, instruction following and safety behavior. The model keeps Google's recommended sampling settings, including thinking mode on, a temperature of 1.0 and a top_p of 0.95, and it runs on the same vLLM serving stack as the original.
Accuracy held up better in some areas than others. Instruction following and agentic tasks gained 0.65 points, and math gained 0.39 points. Code was essentially flat, down 0.02 points. Knowledge and reasoning dropped 0.51 points on average, and long-context and multimodal tasks dropped 0.92 points. The company says most individual results fall within the noise of its tests. Of the twelve benchmarks, five came out ahead of the original model.
The biggest accuracy loss came on AA-LCR, a long-context benchmark, where the score fell 4.00 points. The developer calls it the one real cost in the suite. IFBench, which tests whether a model follows formatting and content constraints, rose 1.83 points, the largest gain. GPQA-Diamond fell 0.96 points and MMMLU fell 0.66 points.
The reductions in thinking were consistent across every benchmark. Long-context and multimodal tasks saw the deepest average cut at 54.4%, followed by knowledge and reasoning at 44.4%, math at 36.1% and code at 27.6%. Instruction following and agentic tasks saw the smallest reduction, at 17.2%. On individual tests, LongBench v2 dropped 69.1%, AA-LCR fell 59.3% and MMLU-Pro fell 49.1%. The smallest cuts came on IFBench at 13.7% and the tau2-bench at 20.8%. The company says those two tasks depend less on long reasoning, since they involve following a list of format rules or applying a fixed customer-service policy through the correct tool calls.
Looping, which the company measured with a compression-based detector that flags text far more repetitive than ordinary prose, also fell. The original model looped on 3.64% of its responses, compared with 0.76% for ThinkingCap. On the three math benchmarks where looping was most common, the rate dropped from 5.5% to 0.9% on AIME 2026, from 8.6% to 1.2% on HMMT November 2025, and from 17.0% to 4.5% on HMMT February 2026.
The company ran both models through the same internal harness, with identical sampling and serving settings, and capped responses at 65,536 tokens. It ran multiple seeds per benchmark, ranging from four seeds on MMLU-Pro and MMMLU to 64 seeds on the AIME and HMMT math sets, and reported accuracy as a mean with a 95% confidence interval. The original model was measured by the developer rather than taken from published figures. Scoring was exact match for most tasks, while LiveCodeBench ran generated code against test cases, and AA-LCR and tau2-bench were judged.
The company also offers a PRO version that it says uses up to 87.6% fewer thinking tokens, or 62% on average. That version is available through an enterprise contact rather than the public release. The base ThinkingCap model is posted on Hugging Face at bottlecapai/ThinkingCap-gemma-4-26B-A4B-it, and the developer invites questions and requests for specific workloads to be tested at hey@bottlecapai.com.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.