AI News Feed
Market watch
Meta Previews New York Times Cooking Integration for AI and Display GlassesNeo4j Makes Case for Shared Knowledge Graphs as AI Agent ContextDigital Wellbeing Beta Points to Flag Trivia Mini-Game and Screen Time Reminder ReworkKFC to Open 'Open House' Test Restaurant in Texas, Adding Breakfast and Table ServiceOkta Uses Dex AI Agent as Customer-Zero Testbed for Identity ControlsApple Maps in iOS 27 Adds Enhanced Flyover, Local Lists and More
Meta Previews New York Times Cooking Integration for AI and Display GlassesNeo4j Makes Case for Shared Knowledge Graphs as AI Agent ContextDigital Wellbeing Beta Points to Flag Trivia Mini-Game and Screen Time Reminder ReworkKFC to Open 'Open House' Test Restaurant in Texas, Adding Breakfast and Table ServiceOkta Uses Dex AI Agent as Customer-Zero Testbed for Identity ControlsApple Maps in iOS 27 Adds Enhanced Flyover, Local Lists and More
Large Language Models

BottleCap AI Releases ThinkingCap-Qwen3.8-27B, Cutting Thinking Tokens 37.2% at 0.86pp Accuracy Cost

BottleCap AI has released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B that cuts average thinking tokens by 37.2% across 12 benchmarks at a 0.86 percentage-point macro-average accuracy cost, with drop-in deployment builds and a gated license.

AI News Feed · 2026-09-25 05:33

The model's main benchmark numbers use reasoning_effort=xhigh, the chat template default. Every benchmark gets shorter, with cuts ranging from 10.7% to 65.5%. Knowledge and multilingual tasks shrink most: MMMLU drops 65.5%, from 1,656 to 571 tokens, and MMLU-Pro drops 57.3%. GPQA-Diamond falls from 12,772 to 7,267 tokens, a 43.1% cut. IFBench thinks 46.4% less with accuracy nearly flat, from 79.75% to 79.71%.

Long-context retrieval improves. AA-LCR accuracy rises 2.25 percentage points, from 81.75% to 84.00%, with 38.6% fewer thinking tokens. LiveCodeBench v6 edges up 0.07 percentage points while thinking 20.3% less. Agentic results hold close to the base: τ²-bench gives up 1.01 percentage points for a 30.9% cut, and Terminal-Bench 2.1 loses 0.56 percentage points, inside its ±4.26 interval, for a 10.7% cut. The most expensive trade is AIME 2026, where accuracy falls 3.85 percentage points, from 98.13% to 94.27%, for 30.2% less thinking.

The 37.2% figure is the mean of the 12 per-benchmark reductions, BottleCap said. Pooled mean thinking tokens fall from 15,735 to 12,144. Under a 16K-token cap per response, ThinkingCap scores higher than the base model; truncated traces fall from 0.51% to 0.34%, and looping falls from 0.06% to 0.05%.

Qwen3.8-27B exposes a reasoning-effort setting, and BottleCap says compression stacks with it. Deltas compared with the base model at xhigh, averaged over 11 benchmarks, show that at medium the base cuts 52.1% of thinking for a 9.16 percentage-point loss, while ThinkingCap cuts 60.2% for a 9.90-point loss. At low, the base cuts 55.4% for a 9.71-point loss, versus 62.3% and 10.79 points for ThinkingCap. With thinking off, ThinkingCap trails the base by 5.7 percentage points. BottleCap recommends xhigh for the best accuracy-to-token balance and says individual thinking modes will get attention in a future release.

The evaluation ran both models through the same harness on one NVIDIA H200 with vLLM 0.29.0. Sampling was identical: temperature 1.0, top_p 0.95, top_k 20 and min_p 0.0. Multi-seed accuracy is the mean with a 95% interval, with seeds ranging from 32 on AIME 2026 to 1 on MMLU-Pro and MMMLU. MMMLU uses a fixed 10,000-question sample, while the other 11 benchmarks run complete sets. MTP speculative decoding with three draft tokens was measured as accuracy-neutral on AIME 2026, accepting 53% of drafted tokens, about 2.6 tokens per step, matching the base model.

The bf16 checkpoint has 28 billion parameters and accepts image and text input. BottleCap publishes five quantized builds: FP8 at 31 GB for vLLM on Hopper and Blackwell; NVFP4 weight-only at 21 GB for vLLM on Hopper with the Marlin kernel and on Blackwell; NVFP4 W4A4 (AWQ) at 23 GB for Blackwell only; GGUF at 16 to 55 GB for llama.cpp, LM Studio and Ollama; and MLX 4-bit DWQ at 21 GB for Apple Silicon Macs with 32 GB. Serving uses the base model's recipe, with --reasoning-parser qwen3 and the qwen3_xml tool-call parser on vLLM, and thinking returns in a separate reasoning field.

The license is PolyForm Small Business 1.0.0 plus a BottleCap personal-use grant, while upstream Qwen materials remain under Apache-2.0. The repository is gated, and commercial use beyond the small-business license requires a BottleCap agreement. Hugging Face lists no inference provider hosting the model yet. It drops in for Qwen3.8-27B on vLLM or SGLang, BottleCap said.

Editor's Summary BottleCap AI's ThinkingCap-Qwen3.8-27B reduces average thinking tokens by 37.2% across 12 benchmarks at a 0.86 percentage-point macro-average accuracy cost, with the largest accuracy loss on AIME 2026 and a gain on long-context AA-LCR. The model is a gated, drop-in fine-tune of Qwen3.8-27B with multiple quantized deployment builds, but commercial use beyond a small-business license requires a separate agreement.