Across twelve benchmarks, accuracy moved by 0.15 percentage points while thinking tokens dropped 38.9%. BottleCap AI calls this post-trained model ThinkingCap. It sits on top of Google's gemma-4-26B-A4B-it, and both the weights and the GGUF quantizations are on Hugging Face, ready to swap in (original announcement).

This is the fifth model in the ThinkingCap series and the first one built on a Google model. The previous four were done for Qwen3.6-27B, Qwen3.8-27B, NVIDIA Nemotron-3.5-Lightning-30B-A3B and Qwen3.8-Flash-Next.

Whose numbers these are

These are the vendor's own measurements, and BottleCap is unusually explicit about it in the announcement. The base model was not quoted from published figures: the team ran it themselves through the same evaluation suite, keeping Google's recommended sampling preset unchanged, with thinking enabled, temperature 1.0, top_p 0.95 and top_k 64. Both sides of the comparison, in other words, ran the same race.

The method is spelled out too. Twelve benchmarks: AIME 2026, two rounds of HMMT, LiveCodeBench v6, GPQA-Diamond, MMLU-Pro, MMMLU, AA-LCR, LongBench v2, RealWorldQA, IFBench and tau2-bench. Random seeds per benchmark range from 4 to 64, with accuracy reported as a mean and a 95% confidence interval. Two failure modes get their own columns: truncation, where a trace hits the generation cap and never closes, and looping, where the model circles the same reasoning forever.

Under that protocol, eight of the twelve gaps fall inside the run-to-run spread, and five of the twelve actually come out ahead of the base model. Only four move beyond their intervals, and they do not all point the same way: IFBench gains 1.83 points, while GPQA-Diamond loses 0.96, MMMLU loses 0.66 and AA-LCR loses 4.00. That last one is the single real cost in the whole set.

Where the saved time went

A 38.9% average cut is not spread evenly. Long-context and multimodal sheds the most at 54.4%, knowledge and reasoning 44.4%, maths 36.1%, code 27.6%, while instruction following and agentic give up only 17.2%. The deepest single cut is LongBench v2 at 69.1%; the shallowest is IFBench at 13.7%.

Two failure rates deserve their own paragraph. Truncation, where the trace hits the 65536-token generation cap and no answer is produced, falls from 4.35% to 0.72%. Looping falls from 3.64% to 0.76%. For anyone running this in production, those two numbers may matter more than the 38.9%.

One detail deserves caution. The worst accuracy regression, AA-LCR, is also the category where thinking was cut hardest. Long-context reasoning dropped 59.3%, and accuracy dropped 4.00 points to pay for it. The thinking that this model saves is largely taken out of long-context capability. If you are about to ship an agent that leans hard on long-document retrieval, that trade may not be the one you want.

What actually moves the bill

Halving token spend only gets you half the story. What usually constrains a reasoning model in production is not FLOPs but the KV cache: every concurrent sequence holds a cache that grows with its length, and once memory runs out, the number of tasks that fit in parallel collapses.

BottleCap ran a throughput test in a separate post. Same RTX PRO 6000, 8 GB of KV cache, vLLM 0.29.0, prefix caching disabled, 64 concurrent requests, 20-minute window, one AIME geometry problem. The original Qwen3.8-27B finished 24 tasks; the ThinkingCap version finished 112, a 4.7x difference. Traces were only 2.3x shorter, but average task parallelism rose from 15.1 to 24.9 and token throughput went from 319 to 511 tok/s. The mechanism is not exotic: shorter traces mean less cache per sequence, so more of them fit on the card, and each finishes sooner, which turns the cache over faster.

To be precise about it, that 4.7x was measured on a different base model, Qwen3.8-27B, not on Gemma-4. But the Gemma-4 base is arguably the more interesting candidate. Third-party review site CrucibleMark tested its Think mode on a DGX Spark: an overall score of 74.94%, a 91.09-second P95 response time, one timeout out of 49 tasks, and a profile described as a batch DevOps specialist, slow but solid, which is exactly the workload shape that leans on the KV cache. According to that review, the model is a MoE with 25.2B total parameters and roughly 3.8B active per token, released under Apache 2.0, with a 262,144-token context window, and GGUF builds are already available on Hugging Face.

Who should switch now, and who should wait

If your service is long-document QA, codebase retrieval or deep research, hold off: those 4 points on AA-LCR will tell you where it hurts first. If the load is batch jobs, background tasks or agent loops, a full order-of-magnitude drop in truncation and looping may be worth more than 0.15 points of accuracy.

Two practical constraints. First, this is a third-party post-training artifact: the base is Apache 2.0, but the derivative weights carry their own terms, so read the model card on Hugging Face before using it commercially. Second, the PRO version, which the vendor claims shaves up to 87.6% of thinking tokens, is available only through a sales email. There are no public weights.

A third point is the one I keep coming back to. Every ThinkingCap number here comes from the vendor's own evaluation harness, and there is no third-party replication yet. "Fewer tokens at unchanged accuracy" has always been the class of claim that most needs independent checking. So run your own A/B on real traffic before treating 38.9% as a fact.

The real point, then, is not the percentage. Reasoning-model cost optimization is moving from "make the model think less" toward "get more work done per card", and this company has already laid out both halves of that argument in two separate posts.