AI Model Releases

AI Model Releases: the most important AI stories, scored for signal and updated continuously.

• Apple trained its own AI model for China with help from Alibaba — Apple has trained a custom LLM for China in partnership with Alibaba — making Apple the first US company to have a proprietary AI model approved in China — with Apple Intelligence set to roll out in the Chinese market within months of an upcoming iOS update.
• Introducing Gemini 3.7 Flash — Google DeepMind officially launched Gemini 3.7 Flash — three weeks after 3.6 Flash — with FrontierCode jumping from 34.4% to 43.6%, DeepSWE from 49.0% to 65.3%, and AutomationBench from 17.0% to 30.4%, at $0.75/1M input tokens locked through year-end.
• Alibaba's Qwen team releases Qwen 3.8 models with open weights under the Apache 2.0 license — Alibaba releases Qwen3.8-27B under Apache 2.0 — a 27-billion-parameter multimodal model outperforming the larger Qwen3.7-Plus on coding tasks, with 262K native context scalable to 1M tokens via YaRN, available now on Hugging Face.
• New benchmark confirms AI models still perform poorly at visual perception — Moonshot AI's PerceptionBench tests visual perception independently of reasoning and knowledge — no frontier model among 16 tested exceeds 60% accuracy, with GPT-5.6 Sol leading at 59.7%, and many assumed reasoning failures traced back to faulty image reading.
• ARENA: Automated Red-Teaming for Large Audio Language Models — ARENA is an automated red-teaming framework for large audio-language models that crafts audio inputs which, combined with safe text queries, induce harmful outputs — achieving attack success rates of 87.9% on Audio Flamingo 3, exposing safety blind spots invisible to text-only testing.
• Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning — A five-month randomized field study across ten Taipei high schools found that LLM-guided adaptive problem sequencing improved Python exam scores by 0.15 standard deviations over fixed curricula — equivalent to 6-9 months of additional schooling — driven primarily by increased student engagement, not content differences.
• Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents — A study of 10,000+ LM agent communities finds collective AI behavior follows statistical mechanics laws: communication improves accuracy on objective questions but causes consistent rightward political drift on subjective ones — with a formal model that predicts individual belief trajectories better than all baselines.
• The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning — A new multimodal benchmark asks models to infer words from pen-scratch audio and hand-movement video alone — humans score over 80%, while GPT-4o and Gemini 2.5 Pro both fail to surpass 10%, revealing a fundamental gap in cross-modal causal reasoning.
• Self-Supervised Visual On-Policy Distillation — S²VOPD pushes Qwen3.5-4B from 70.7% to 77.4% on six fine-grained visual benchmarks — surpassing GPT-5.4 — using only asymmetric augmented views during training, no privileged annotations, stronger teacher, or reward signals required.
• Beyond Visual CoT: Internalized Visual Thinking for Proactive Video Reasoning — Internalized Visual Thinking (IVT) matches Visual Chain-of-Thought accuracy on video reasoning while cutting inference latency by more than 5x — by training models to predict future-frame embeddings internally rather than generating explicit intermediate images.
• PertMind: Eliciting Emergent Biological Reasoning in LLM via Reinforcement Learning on Cellular Perturbation Data — PertMind trains LLMs on cellular perturbation prediction using gene-response data as RL rewards — no manual reasoning traces required — and the resulting models transfer without task-specific fine-tuning to reverse perturbation identification, double-perturbation reasoning, and phenotypic screen prioritisation.
• Training-Free Knowledge Transfer Across Model Scales through Activation-Guided Pruning — Activation-Prune-Merge (APM) transfers capability from a large donor LLM to a smaller 3B recipient without any retraining, lifting average accuracy from 55.5% to 60.6% across 16 benchmarks — with RTE jumping from 64.3% to 82.3%.
• Bounded Semantic Planning and Deterministic Compilation for Reliable Enterprise Text-to-SQL — Semantic Path Compilation (SPC) achieves 97.4% accuracy on an insurance enterprise SQL benchmark versus 55.3% for direct LLM-to-SQL, by separating stochastic intent interpretation from deterministic query construction.
• LORA-CRAFT: Cross-layer Rank Adaptation via Frozen Tucker Decomposition of Pre-trained Attention Weights — LORA-CRAFT (CRAFT) achieves LoRA-level or better accuracy on LLaMA2-7B and LLaMA3-8B using hundreds of times fewer trainable parameters, by applying Tucker tensor decomposition to pre-trained attention weights across layers and training only small adaptation matrices on frozen factors.
• Rethinking Automated Program Repair: The Impact of Bug Complexity, Fault Localization, and LLM Cost-efficiency — An ESEM 2026 empirical study finds over 50% of moderately complex bugs can be fixed by low-cost LLMs, DeepSeek-V3.2 offers the best cost-efficiency, and GPT-5 repairs 39 more complex bugs than DeepSeek-V3.2 — though imprecise fault localization widens performance gaps more than model choice.
• Counting Documents Is Not Counting Text: Unit Bias in Web-PDF Corpus Statistics — Analysis of 7.9M web PDFs shows Common Crawl's truncation cap affects 63% of corpus text — not just 23% as document-count statistics suggest — and two widely used recovery libraries reclaim only 11.4% and 1.4% of that text, meaning 55-62% of total corpus tokens are silently lost.
• LAVA: Logic-Aware Validation and Augmentation Framework for Large-Scale Financial Document Auditing — LAVA, a modular multimodal LLM pipeline for financial document validation, outperforms baselines on hallucination control and edge-case handling across real-world payroll, tax, and loan underwriting documents — with traceable symbolic verification of arithmetic rules.
• CACSurv: Concordance-Aligned Comparative Learning with Large Language Models for Cancer Survival Prediction — CACSurv reformulates cancer survival prediction as LLM comparative ranking across patient mini-cohorts, achieving a C-index of 0.722 on six TCGA cancer types — 6.5 percentage points above the best prior survival model and 4.2 pp above LLM time-regression baselines.
• T-LLM Compiler: Trusted LLM-based Code Optimization and Verification Framework — T-LLM Compiler combines LLM-generated code transformations with traditional compiler verification in an iterative feedback loop, achieving 83.3% optimization accuracy and up to 16.1% runtime speedup on PolyBench/C benchmarks (26.7% average speedup vs standard baselines). Source code is open-sourced.
• ReasonCast: Agentic Demand Forecasting with Selective Semantic Reasoning — ReasonCast reduces demand forecast error by 3.29 WMAPE points on holiday-sensitive categories by teaching LLMs to selectively intervene with structured event fields (direction, shape, amplitude, peak) — while suppressing unnecessary intervention on stable periods to avoid a 1.68-point penalty.
• When Context Misleads: Intent-Guided Decoding for Robust Retrieval-Augmented Generation — Intent-Guided Decoding (IGD) fixes a core RAG failure mode — blindly trusting retrieved context — by arbitrating between external evidence and model memory based on user intent, recovering up to 65.4 percentage points on factual-conflict benchmarks.
• MobileMem: Learning from a Year of Mobile Experiences — MobileMem introduces a year-scale mobile experience benchmark for on-device long-term memory in AI agents, covering multi-hop reasoning, temporal reasoning, knowledge updating, and implicit preference inference across text and multimodal settings.
• Think in Latent, Explain in Language: Self-Explainable Latent Reasoning — Self-Explainable Latent Reasoning (SELR) trains a single model to reason efficiently in latent space while decoding its own latent representations into human-readable Chain-of-Thought steps—outperforming both efficiency-only (Coconut) and separate-decoder (Heima) approaches on accuracy, token efficiency, and interpretability.
• OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing — Researchers release OpenGPT-4o-Image, an 80K instruction-image pair dataset spanning 11 domains and 51 subtasks including chemistry diagrams and complex multi-step editing — fine-tuning on it improves image editing benchmarks by up to 18% and generation by 13%.
• What the Reranker Sees: Multi-Aspect Page Annotation for Long-Document Multimodal Question Answering — Trident improves long-document VQA by converting retrieved pages into structured semantic records (visual captions, entity tags, section paths) before reranking — adding up to 6.6 accuracy points on synthesis questions where current retrievers like ColPali reach only F1@5=0.332.