# ThinkingCap ThinkingCap is a series of open-weight models from BottleCap AI, post-trained so that they reach the same answer with a shorter thinking trace. Their own goal statement: "spend fewer tokens getting to the answer, and change nothing else." Why it matters: [[AI Reasoning Models]] are billed by the token and slowed down by the token. A reasoning model that thinks 40% less is 40% cheaper and faster on the thinking part, IF the answers stay as good. That's the bet. The fifth model in the series, released on 23 October 2026, is built on Google's [[Gemma 4]] 26B-A4B (a [[AI Mixture of Experts (MoE)|mixture of experts]] with about 4B active parameters per token). Earlier ThinkingCap versions exist for Qwen models. ## Results on Gemma 4 26B-A4B BottleCap's own evaluation, thinking on, Google's recommended sampling settings (temperature 1.0, top_p 0.95, top_k 64), served with [[vLLM]]: - **38.9% fewer thinking tokens** on average across 12 benchmarks (8,773 → 5,398 tokens per answer) - **Up to 69.1% fewer** on LongBench v2, and only 13.7% fewer on IFBench, where the tasks need less reasoning in the first place - **Average accuracy: 75.18% → 75.03%** (-0.15 points). Most differences sit inside run-to-run noise - **Fewer runaway traces**: looping fell from 3.64% to 0.76% of responses, and truncated traces (the model never closes its thinking) from 4.35% to 0.72% A few rows from their table: | Benchmark | Accuracy base → ThinkingCap | Thinking tokens | |---|---|---| | AIME 2026 | 86.20 → 86.72 | -39.3% | | GPQA Diamond | 79.92 → 78.96 | -44.4% | | LiveCodeBench | 76.98 → 76.96 | -27.6% | | MMLU-Pro | 84.55 → 84.64 | -49.1% | | AA-LCR (long context) | 64.62 → 60.62 | -59.3% | | LongBench v2 | 59.43 → 60.88 | -69.1% | | τ²-bench (agentic) | 64.25 → 63.73 | -20.8% | AA-LCR is the one real loss: 4 points of accuracy on long-context reasoning. ## How it's done BottleCap doesn't say. The post and the model card only mention "shorter-thinking post-training" ([[AI Post-Training]]), with no algorithm, reward or data. The obvious guess, also made on Hacker News, is reinforcement learning with a penalty on trace length. There's also a paid "PRO" version claiming up to 87.6% fewer thinking tokens (62% on average), for enterprise customers. ## Running it - Weights on [[HuggingFace]]: `bottlecapai/ThinkingCap-gemma-4-26B-A4B-it` (BF16, about 52 GB) and a GGUF build (MXFP4_MOE, 16.6 GB, with the vision projector) for [[llama.cpp]], [[Ollama]] and similar tools - Apache 2.0, the base model's license - Drop-in replacement for the base model on the same stack. Thinking is off unless the request sets `enable_thinking=True` - vLLM: `vllm serve bottlecapai/ThinkingCap-gemma-4-26B-A4B-it --max-model-len 262144 --reasoning-parser gemma4 --enable-auto-tool-choice --tool-call-parser gemma4` - [[Docker Model Runner]]: `docker model run hf.co/bottlecapai/ThinkingCap-gemma-4-26B-A4B-it` ## What Hacker News thought The thread (127 points) was lukewarm: - Several people found reduced-thinking models weaker on their own tasks than on benchmarks. One said ThinkingCap Qwen was "slightly dumber" than the original and refused more. Another reported that a similar trimmed model from [[Fireworks AI]] (Ember-1, based on [[Kimi K3]]) hallucinated and claimed tasks were done when they weren't - Many said you get most of the speed by setting the reasoning effort to low or medium on the base model, with no fine-tune at all. [[Simon Willison]] noted that [[Qwen3.8-27B]] defaults to maximum thinking and gets much faster at medium or low - Others argued that faster inference engines (1.5x or more in tokens per second) matter more than fewer tokens, and that wall-clock time per correct answer is the metric that counts - One commenter pointed to a use case where it fits: PR reviews, where a shorter trace at equal accuracy beats a slightly higher score ## My take The idea is right: a lot of thinking tokens are waste, and loops and never-ending traces are real failures, not just cost. Cutting looping from 3.6% to 0.8% may be the most useful number in the whole post. But I'd treat it as a promising benchmark result, not a proven model. The evaluation is BottleCap's own harness, the method is secret, and the HN reports point to the usual problem: models trained to think less lose something that benchmarks don't measure. Before reaching for a fine-tune, I'd first try the base model at a lower reasoning effort. If that's not enough and you run Gemma 4 locally, ThinkingCap is cheap to test: same stack, same license, one model name to change. ## References - BottleCap AI, "ThinkingCap: Gemma-4-26B-A4B" (2026-10-23): https://bottlecapai.com/post/thinkingcap-gemma-4-26b/ - Model card: https://huggingface.co/bottlecapai/ThinkingCap-gemma-4-26B-A4B-it - GGUF: https://huggingface.co/bottlecapai/ThinkingCap-gemma-4-26B-A4B-it-GGUF - Base model: https://huggingface.co/google/gemma-4-26B-A4B-it - Hacker News discussion: https://news.ycombinator.com/item?id=50209907 ## Related - [[Gemma 4]] - [[AI Reasoning Models]] - [[Chain-of-Thought (CoT) prompting]] - [[Latent Reasoning]] - [[AI Post-Training]] - [[AI Cost Management]] - [[AI Inference]] - [[Running AI Models Locally]] - [[AI Open Weight Models]] - [[Claude Code effort levels]]