Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 3 min read

Hugging Face AI Updates: November 11, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Hugging Face Documents Decision Models, Which Return Probabilities Over Fixed Answers, With Transformers DecisionPipeline and llama serve Support

Hugging Face. A November 10 blog post introduces decision models, which take a state and a set of questions with predefined answers and return structured decisions with probabilities instead of generating text. It defines three question types: choice (a softmax over options), score (an ordered scale reported as an expected level) and yes/no (the probability of “yes”). Transformers gains DecisionPipeline, AutoModelForDecision and DecisionConfig, used as pipeline("decision", model="autotrust/GEV-26B-Decide"). Serving is available through a /v1/systemone endpoint in transformers serve and the same API in llama.cpp, the Hub has a decision-model tag, and fine-tuning works with the Trainer API. TRL support (PR #8040) was still open when the post was written, and an Unsloth recipe covers LoRA-based decision heads. The post’s worked confidence example uses two different values for the same probability. Source

2. Hugging Face Hub Organizations Can Now Be Billed Through Google Cloud Marketplace

Hugging Face. A November 10 post describes how an organization admin can link a Hugging Face organization to a Google Cloud account through the Google Cloud Marketplace, so usage appears on the existing Google Cloud invoice instead of a credit card. It covers Team or Enterprise subscriptions, pay-as-you-go compute (Jobs, Inference Endpoints, Inference Providers, GPU Spaces and ZeroGPU quota extensions) and extra storage, all at public Hugging Face pricing. Hugging Face reviews each subscription before marketplace billing starts, and organizations with a card on file switch at the start of their next billing period. Included monthly plan credits are applied before usage is reported to Google Cloud, adding credits is unavailable while billed this way, and linking and unlinking are recorded in audit logs. Source

3. Tether’s BotPsy Is a 0.8B Agentic Model That Raises BFCL v4 From 23.4 to 44.0 Under Apache 2.0

Hugging Face / Tether AI Research. Tether AI Research published BotPsy on the Hugging Face blog on November 11, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows on personal devices. Against its base model it raises BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, tau-bench from 23.2 to 51.5 and IFEval from 59.7 to 90.8, and it leads Qwen3.5-2B-Think on five of six tool-use and agentic benchmarks, with Claw-Eval the exception (46.5 against 48.9). Training used four stages: supervised fine-tuning, multi-teacher on-policy distillation and two reinforcement learning rounds. Automated red-teaming found unsafe responses 21% of the time versus 36% for the base model. Weights are Apache 2.0, with safetensors and GGUF files (bf16, Q8, Q4) for the QVAC SDK and llama.cpp described as forthcoming. The authors say it suits low to medium difficulty tasks and is not intended for high-stakes decisions. Source