📁 LLM

HauhauCS releases two uncensored, balanced Gemma 4 variants with QAT 4-bit quantization and Multi-Token Prediction (MTP) for speculative decoding, yielding up to 53% speed gains without quality loss on consumer hardware. The models, sized 16.8 to 18.7 GB VRAM in Q4_K_M, target on-premise control and data sovereignty.

Google has made computer use a built-in feature in Gemini 3.5 Flash, removing the need for a standalone model. This simplification speeds up agentic AI deployment, but enterprise trust hinges on control, transparency, and data sovereignty—key concerns for organisations evaluating such capabilities.

Meta has begun testing an app with a built-in AI assistant for creators. The rollout raises questions about processing location—cloud or local—and its impact on data sovereignty, a key factor for those considering on-premise deployment.

Switzerland's highest court is evaluating an “abliterated” language model to circumvent unwarranted refusals from standard systems. A concrete case raising questions about alignment, on-premise control, and data sovereignty in the legal use of AI.

The startup builds software that shrinks AI models while preserving performance, targeting on-premise and edge deployments and cutting carbon emissions.

Baidu releases Unlimited-OCR on ModelScope: 3.3 billion parameters, MIT license, one-shot parsing of images, PDFs, and multi-page documents. 32K output length, Transformers inference and SGLang serving with OpenAI-compatible streaming. A building block for on-premise OCR without cloud dependencies, handling complex layouts. The full-document approach and extended context window target enterprise scenarios with privacy requirements.

Qwen has released AgentWorld-35B-A3B, a 35B-parameter MoE with only 3B active per token. It's not a chatbot but a world model designed to predict how seven interaction domains — terminal, Android, web, OS GUI, and more — respond after an agent action. A resource for training, testing, and evaluating agents offline, without running actual tools.

A new reinforcement learning approach assigns fine-grained rewards to individual SQL clauses, improving the accuracy of Text-to-SQL models. Concrete implications for those running inference on-premise with proprietary databases.

A study training six offline reasoning methods on Qwen3-4B finds that SFT, RFT, and RIFT produce nearly identical weight updates, while DPO diverges sharply and achieves the highest accuracy. A geometric analysis useful for those choosing fine-tuning strategies for on-premise infrastructure.

Researchers used reasoning traces from classical rule-based planners to supervise a small 4B-parameter driving VLA, achieving significant reductions in trajectory error and miss rate. The method ensures that reasoning is causally tied to motion planning, a key point for those considering compact models for on-premise deployments.

ByteDance's new video model, unveiled in Beijing, generates 30-second native 4K clips while accepting up to 50 reference inputs. A four-version skip signals a generational leap, with an enterprise beta already active. For those evaluating on-premise deployment, open questions remain about hardware requirements and data sovereignty.

Anthropic has launched Claude Tag in research preview, an integration of Claude with Slack that lets users tag @Claude for insights and task assignment. Available to Enterprise and Team customers, the feature points to a future of persistent AI assistants in work tools. But its cloud-native nature reignites the debate over data sovereignty and on-premise alternatives.

A paper shared on Hugging Face provides new evidence but not definitive proof. For those running LLMs on-premise, this nuance is critical: it shows that every claim must be verified in one's own stack, because reproducibility and data security rely on real-world tests, not just published research.

Anthropic has introduced Claude Tag, a new feature aimed at organizing and managing interactions with its LLM models. For those operating on-premise, tagging tools can strengthen data governance and regulatory compliance. AI-RADAR examines the implications of this move, while noting that technical details remain scarce.

Immunologist Derya Unutmaz cracked a three-year mystery about T cell behavior using GPT-5 Pro. The model spotted patterns that traditional analysis missed, potentially advancing cancer and autoimmune therapies. The case reignites the debate on integrating large language models into biomedical research, balancing compute power, data privacy, and architectural choices.

Omio integrates ChatGPT and Codex across engineering, reducing development effort to 20% and compressing timelines. A conversational booking interface grounded in live data marks a shift to conversational commerce. But governance remains key: humans retain full accountability, with AI as a speed booster.

The Krea 2 Turbo model is now available for download on Hugging Face. The 'Turbo' label suggests optimizations for low latency and reduced VRAM usage, a signal for those considering on-premise deployment who want to maintain data control without sacrificing speed.

Anthropic pinpointed a fix for spikes in errors across multiple Claude models, while still explaining why Claude Mythos 5 and Claude Fable 5 were suspended. The incident reignites debate about cloud LLM reliability and the control on-premise can offer.

A new micro-benchmark evaluates Large Language Models on writing datafiles for Surface Evolver, a 1992 tool for solid-liquid interfaces. With 8 rounds of autonomous debugging, it provides objective scoring and challenges models on sparse-training scientific tasks – a useful angle for those selecting LLMs in on-premise settings.

Some users report that GLM 5.2 stands out for its blunt, no-fluff attitude, avoiding the sycophantic tendencies of many US models. This difference may stem from culturally-informed training data, with implications for on-premise LLM selection when organizational values and directness are priorities.