ThinkingCap-Qwen3.8-27B is a fine-tune of Qwen3.8-27B focused on shorter reasoning traces. It reduces average reasoning tokens by 37.2% across 12 benchmarks while lowering macro accuracy by 0.86 percentage points and improving some long-context tasks.
With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Under a 16K-token cap per response, ThinkingCap scores higher than the base model.
BottleCap AI Releases ThinkingCap-Qwen3.8-27B: 37.2% Fewer Thinking Tokens at a 0.86pp Accuracy Cost.
# What ThinkingCap-Qwen3.8-27B Does
BottleCap AI fine-tuned Qwen3.8-27B to produce shorter reasoning traces without changing answer style or adding knowledge. The explicit goal: reduce unnecessary "thinking" tokens that the model emits during multi-step reasoning while keeping instruction following, safety behavior, and core reasoning ability largely intact.
# Main results at a glance
# Benchmark-level trade-offs
# Effort dial interaction
Qwen3.8-27B exposes a reasoning-effort setting. BottleCap reports that ThinkingCap's compression stacks with that dial. Compared to the base model at xhigh, ThinkingCap cuts thinking tokens more aggressively across medium and low effort settings while keeping accuracy shifts similar. With thinking disabled altogether, ThinkingCap trails the base by 5.7pp. BottleCap recommends xhigh for the best accuracy-to-token balance and plans future work on per-mode tuning.
# Evaluation setup and reproducibility
# Deployment and builds
The bf16 checkpoint contains 28B parameters and accepts image and text inputs. BottleCap provides five quantized builds:
Serving follows the base model recipe with --reasoning-parser qwen3 and the qwen3_xml tool-call parser on vLLM. Thinking outputs appear in a separate reasoning field.
# License and access
Weights are gated. The default license is PolyForm Small Business 1.0.0 plus a BottleCap personal-use grant. Upstream Qwen materials remain under Apache-2.0. Commercial use beyond the small-business license requires a BottleCap agreement. No inference-provider hosting was listed on Hugging Face at the time of reporting.
# Practical implications
If you need to reduce token consumption for cost or latency reasons while accepting small accuracy losses on average, ThinkingCap offers a straightforward substitute for Qwen3.8-27B on vLLM and SGLang. Task-specific evaluation is necessary: some benchmarks maintain accuracy with large token savings, while a few (notably AIME 2026) show meaningful accuracy drops.
Qwen 3.8 27B introduces a notable enhancement to text generation workflows through its Multi-Token Prediction (MTP) feature. This capability enables the model to predict multiple tokens in a single step, significantly increasing processing speed without compromising output quality. With MTP enabled, users can achieve s
Fireworks AI has released Ember-1, a post-trained Kimi K3 that learns to produce shorter reasoning traces instead of lowering reasoning effort. Fireworks reports about 40% fewer tokens, with output tokens per task falling from 49.3K to 29.9K in a production A/B test at an essentially unchanged score. Ember-1 is availab
Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, the same model behind two different safeguard layers. Fable 5.1 is generally available on the Claude API, AWS, Google Cloud, and Microsoft Foundry; Mythos 5.1 remains restricted to vetted organizations under Project Glasswing. Fable 5.1 scores 52.6% on Term
Alibaba's Qwen3.8-Omni-Flash understands audio and video, plans tasks, calls tools, and reports about 45.7% fewer tokens on OmniVideoBench. The post Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use appeared first on MarkTechPost .
Your customers expect consistent, instant, intelligent service across every channel, even when they bounce between SMS, email, and web chat from one interaction to the next. But your team can’t be everywhere at once—no one can. That’s exactly where agentic AI in retail comes in. Agentic AI gives customers a seamless ex
Agentic AI has the potential to change SaaS from software that users primarily operate into software that can increasingly perform bounded business outcomes on their behalf. That is a much more significant change than adding an AI assistant to an existing application. The transition affects product architecture, user e
Loading more related stories...
Open the app view to save this story, compare related coverage, and continue from the same source.