Latest in AI
ThinkingCap Gemma 4 26B: Fewer Reasoning Tokens, Benchmarks and Deployment
BottleCap’s new Gemma derivative shortens reasoning traces, but the average hides a tradeoff worth checking before replacing your model.
BottleCap AI released ThinkingCap Gemma 4 26B on October 23, a post-trained version of Google’s Gemma-4-26B-A4B-it. The company reports 38.9% fewer thinking tokens across twelve benchmarks, with average accuracy falling 0.15 percentage points. It gives self-hosting teams another way to test reasoning efficiency without changing model families.
What ThinkingCap changes
This is BottleCap’s derivative, rather than a new Google release. The available checkpoint supports image input and uses the base model’s Apache 2.0 license. BF16 weights occupy about 52 GB. That is weight storage, not the total memory required to serve requests.
Thinking is off by default. The reported comparison turns it on, so loading the weights without enabling reasoning will not reproduce the advertised setup. The model card documents Transformers and vLLM usage, including the per-request enable_thinking setting.
The practical question is whether a shorter reasoning trace still does the work your application needs. A concise answer can look perfectly normal even when the model skipped a useful intermediate check. Evaluate the outcome, not how convincing the explanation sounds.
ThinkingCap Gemma benchmarks
BottleCap’s release report compares the models on the same serving stack and sampling settings, using multiple seeds. These are the developer’s results, not an independent replication.
| Test | Base accuracy | ThinkingCap accuracy | Thinking-token reduction |
|---|---|---|---|
| LiveCodeBench | 76.98% | 76.96% | 27.6% |
| AA-LCR | 64.62% | 60.62% | 59.3% |
| LongBench v2 | 59.43% | 60.88% | 69.1% |
The headline 38.9% is the mean of twelve per-benchmark reductions. It is not a pooled workload saving. Average accuracy moves from 75.18% to 75.03%, while AA-LCR loses four percentage points. A small average change can hide a meaningful loss on the task you care about.
For a coding assistant, the coding row is a more relevant starting point than the largest percentage in the table. For document analysis, inspect the long-context tests separately rather than treating them as interchangeable evidence.
How to evaluate the release fairly
The model card adds an important boundary: LongBench v2 covers 370 of 503 questions whose prompts fit within 196,000 tokens. Its published results therefore should not be read as a full-suite score. The evaluation also uses BF16 models, so it does not establish that every quantized build preserves the same accuracy.
A useful comparison keeps the prompt, tools, generation limit and sampling settings fixed. Then measure successful task completion, total generated tokens and time to completion. If one model needs a retry or an extra tool call, include that work rather than stopping the measurement at its first response.
This matters particularly for agents. A cheaper individual model call can still produce a more expensive workflow if the agent makes a weaker decision and has to recover. Conversely, avoiding an unnecessarily long reasoning trace may make a repeated background task easier to run within a fixed time budget.
Downloads, pricing and deployment limits
The release includes a GGUF build for smaller deployments. Its listed MXFP4_MOE file is 16.6 GB, with a separate vision projector. Leave room beyond the weight files for context and runtime overhead when assessing hardware requirements.
BottleCap also advertises an enterprise PRO variant through sales contact. The announcement does not give a public per-token price. The downloadable checkpoint’s economics depend on how you host and use it.
Fewer thinking tokens do not imply the same percentage reduction in latency or total cost. Input processing, final-answer tokens, memory and the rest of the application still contribute. Measure those on the hardware and traffic pattern you intend to use.
ThinkingCap is most useful as a controlled substitution experiment for teams already considering this Gemma model. The interesting next result is whether its shorter traces preserve your own acceptance criteria, not whether the average reduction survives being repeated in another headline.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn model releases and engineering updates into clear stories that developers and founders want to read and share.