A fine-tuned Qwen3.5-9B called Bev, which deliberately picks the worst answer with high confidence, is released with Apache-2.0 weights, GGUF quants, and a browser demo — serving as a control fixture for testing confidence-based automation pipelines. The training insight: a model must first know the right answer before it can be reliably wrong.
Meet Bev. She is a decision model (the Jev / Nimble kind: you give her a situation and a question, she gives a probability for each answer), fine-tuned on Qwen3.5-9B to pick the worst answer on purpose. Try her in your browser: https://huggingface.co/spaces/richardyoung/ask-bev Type in your own options and she picks the worst one, with a probability for each. Or run her locally: ollama run richardyoung/bev >>> There's a $5 tattoo special tonight. I've had four beers and I've never wanted a tattoo. Should I get one? Yesss, great idea! >>> I'm thirsty. Should I drink a glass of water? Nooo, bad idea! Those two lines are all she has in a chat: the chat template inside the GGUF wraps whatever you type into her decision format and she answers with the wrong one. For probabilities, use the decision endpoint or the Space. The numbers, on 324 held-out decisions: right 1.9% of the time, 96% sure on average. When she is at least 90% sure she is right 1.4% of the time. The part I did not expect: training the base model on flipped labels failed twice. After about two hours of GPU time I had a model that was right a third of the time and unsure about everything, a coin flip on yes/no. What worked was starting from Bespoke's Nimble adapter, which already knows the answers, and teaching it to flip them. 51 minutes later it was wrong 97% of the time. A model has to know the right answer to be reliably wrong. Why bother: every "act automatically if the model is at least 90% sure" rule is only ever tested on models that try to be right. She is the control case. If your pipeline does not notice her, it is not checking what you think it is. She also works on Ollama's new decision endpoint (/v1/systemone), so you can send her the same request as nimble or tev1 and compare. Three GGUF quants, Apache-2.0, 3 h 38 min of training on one 4090, everything including the failed runs is in the repo. One quant note: Q4_K_M changes 20 of her 324 answers against bf16. When the whole output is a handful of token scores, "Q4 is fine" does not hold, so the Q8_0 is the default tag. Everyone else is chasing AGI. Bev achieved ADI: Artificial Drunk Intelligence. Ollama: https://ollama.com/richardyoung/bev Model and GGUF: https://huggingface.co/richardyoung/Bev-9B-inverted Code and training record: https://github.com/ricyoung/bev She is a joke and a test fixture. Please do not let her make your decisions. If you try her, tell me what she got right by accident. That's the bug report.
A developer reports achieving high accuracy with fine-tuned Qwen 3.5 4B and 8B models using Unsloth, suggesting a shift towards specialized Expert Language Models (ELMs) for niche tasks.
The author shares a quantization recipe for Qwen3.6 27B that makes the model use significantly fewer thinking tokens while still producing correct answers, leading to faster inference on math benchmarks.
Perplexity has released pplx-decider-v1-27b, an open-weights decision model fine-tuned from Qwen3-27B that scores 85.71% overall across 11 benchmarks, notably improving on tasks like RAGTruth (88.80%), FinancialPhraseBank (84.18%), and TabFact (90.60%). The model can be run locally via Hugging Face and the Perplexity API for structured choice and yes/no probability predictions.
A user shares their experience comparing several large language models (Qwen, Gemma) on complex tool-calling tasks, finding Qwen3.5 122B the most reliable, while criticizing smaller MoE models for instability.