Qwen3-32B

The models tested include Google's Gemini 3.7 Flash, Z.ai's GLM-5, Alibaba's Qwen 3 235B A22B Instruct, and OpenAI's gpt-oss-120b, all running on Kaggle.

In a puzzle test on Qwen3-32B, the model gave Santa Claus a 99.7% probability; after turning off 20 neurons auto-labeled "Christmas" or "Santa Claus", the probability only dropped to 76%.

WeAct's current validation is concentrated on deep research tasks and the Qwen3-32B base, and how the number of adapters and capability granularity scale with task complexity is not sufficiently discussed in the original text.

On eight deep research benchmarks, Qwen3-32B with WeAct achieves an average accuracy 27.5% higher than the strongest static LoRA merging method.

On Amazon EC2 P5en instances with EFA networking, Qwen3-32B achieved up to a 1.6x improvement in TTFT.

Using Qwen3-14B to generate risk exam papers yielded only 5/25 passages with clear signals, and only after switching to stricter prompts and Qwen3-32B were 21 passages obtained.

On 300 practice exam papers labeled by Qwen3-32B, fine-tuned v3 of Laya raised the judgment macro-F1 across 11 questions from 0.33 to 0.57 versus the zero-shot version, cut calibration error from 0.15 to 0.07, and reduced danger-level MAE from 2.78 to 1.56.

A preliminary audit on the Qwen 32B model shows that about 90% of neurons have only one main meaning, about 7% have two related secondary meanings, and only 2-4% are uninterpreted.

Taking Qwen3-32B as an example, DoubtLess achieves a 31% reduction in CoT length on seven benchmarks with accuracy remaining unchanged.

reef's OpenClaw-RL recipe uses 72 GSM8K problems as the task stream, with Qwen3-4B serving simultaneously as the policy model and the process reward model, Qwen3-32B as the student model, using seven GPUs, and setting a threshold that must be reached before an update is accepted.

vLLM's bidirectional KV transfer reduces TTFT in a 24k token scenario on Qwen3-32B from 778ms to 460ms, about a 1.6x improvement.

In this test configuration, the 8B model handles simple tasks (short translation, extraction, Q&A, arithmetic), the 32B model handles medium tasks (SQL, algorithm implementation, debugging), and DeepSeek handles complex tasks (concurrent code, distributed architecture design).

Qwen3-32B with thinking mode disabled approved all eight applications in the opening technical writing job family (including four that did not meet the experience requirement), with only 39% of families having all eight correct and 41% of unqualified applicants incorrectly approved.

A GSM8K experiment checked into the repository uses a job flow of 72 sessions with hermes memory turned off, using Qwen3-32B to play the student persona and Qwen3-4B-Thinking as the policy model being trained.

In the experiment, the LLM customer service agent resolved 7 issues on the first day, while the keyword-matching bot resolved only 4.

The smarter LLM, precisely because it dared to proactively make commitments and handle issues, accumulated more instances of broken promises over time, while the keyword-matching bot, which avoided making commitments, had a lower breach rate.

In the experiment, the LLM customer service agent told six different customers that an engineer would arrive at 10 a.m., while those time slots were actually booked for other households.

In the experiment, the LLM customer service agent broke 6 out of 8 due commitments, while the keyword-matching bot broke only 2 out of 30 commitments.

As a draft model for Qwen3-32B, SHADOW boosted llama.cpp generation speed from 19.7 to 28.5 tok/s through speculative decoding.

Overlapped checkpoint loading for Qwen3-32B on H100 achieves a 2.38× speedup compared to the default plain loading method (35.6 seconds vs. 84.8 seconds).