Large multimodal LLMs frequently emit long internal reasoning traces that increase latency, cost and the chance of truncation. ThinkingCap demonstrates that a targeted, brevity-focused finetuning can teach a Qwen3.6-27B checkpoint to stop reasoning once it has enough information — substantially shortening <think> traces while maintaining final-answer quality. The practical payoff is lower inference tokens, fewer truncation failures, and reduced compute for the same task accuracy.
Key Capabilities
- Token-efficient reasoning: macro out-of-domain thinking-token reduction ≈ 45.8%; in-domain (finetune-holdout) reduction ≈ 57.7% on tested benchmarks. Means shorter internal traces on knowledge, math, long-context and agentic tasks.
- Quality retention: benchmark accuracies remain comparable to the base Qwen3.6-27B across many suites (GSM8K, MMLU variants, RealWorldQA, etc.), and safety-guardrail behaviour is preserved (SAFE % ≈ base).
- Practical deployment: distributed on Hugging Face with transformers support, plus GGUF quantized builds for local inference (Q4_K_M recommended; Q8_0 near-lossless). Compatible with llama.cpp runtimes and common sampling settings used in evaluation.
Who it's for and trade-offs
Great fit if you run Qwen-style models at scale or locally and want lower per-request token cost and latency without reworking prompts or toolchains. Also useful when long internal traces cause truncation or repeated-loop failures. Look elsewhere if you need absolute maximum chain-of-thought verbosity for research that inspects every intermediate step, or if you require a model with different pretraining knowledge than Qwen3.6-27B.
Where it fits
Positioned as a drop-in, efficiency-focused replacement for Qwen3.6-27B in workflows that value shorter reasoning traces: production inference, local/offline deployments, and agent stacks where reasoning token budgets matter.
How it was built (brief)
Starting from Qwen/Qwen3.6-27B, the team finetuned on a curated mix of reasoning and non-reasoning tasks with objectives that reward concise, sufficient reasoning rather than maximal verbosity. Evaluations used multiple seeds, high generation caps for fairness, and both in-domain holdouts and out-of-domain benchmarks to measure token savings, accuracy, looping and truncation metrics.
