Google Research details a 'sleep' phase for deployed LLMs: unlock spare MoE experts, distill short-term context into them, then self-train on filtered synthetic 'dreams'. Google reports near-perfect BABILong scaling to 10M tokens. research.google/blog/the-sle…

Liquid AI's LFM3 swaps LFM2's hybrid stack for a tri-hybrid mix of convolutions, sliding-window and global attention, trained on 22T tokens. Six ~2B models share one backbone with 256K context. Liquid reports 1.7B beats LFM2.5-2.6B at 35% fewer parameters. liquid.ai/blog/lfm3

Google DeepMind's Genie 4 is its first real-time world model to generate audio natively: ambient sound, music and dialogue come with the world. A Gemini 3.8 Flash agent plans steps and checks frames as it runs. Preview now; SynthID on audio and video. deepmind.google/blog/genie-4…

Opera started shipping local AI models in its browser. WebAI, which Opera says Google and Microsoft back, is a proposed standard in early preview in Opera One. It runs Google's Gemma and, on Macs, Apple's on-device model. Opt-in, per-site permissions. next.opera.com/blog/web-ai-l…

Mirendil released Fern 0.1, a from-scratch coding model with a 10M-token context. Its base holds 100% on RULER single-needle retrieval to 10M, but Fern 0.1 scores 6.1% on Terminal Bench 4.0 vs 66.4% for Claude Opus 5.5. Early access only. mirendil.com/news/introducin…

Google researchers and academic collaborators posted StreamVibe, a streaming video agent that decides for itself when to speak, even when a request never names a trigger. They report 10.7% better response timing and answer quality than baselines. arxiv.org/abs/2610.34662

vLLM Semantic Router released d3-ultra, a 26B open multimodal decision model. It reads an image, a video or text once and returns a probability for every question, no text generated, 131k tokens per question. Apache 2.0. 87.2 on its own SR-Text test. huggingface.co/vllm-sr/d3-ul…

Baidu released Unlimited OCR 2, an MIT-licensed 3B parser that reads 100 to 300 page documents in one request. Its sliding-window attention keeps per-step cost flat, 7x faster than full attention. Baidu reports 77.1 on its own new MarathonDocBench. huggingface.co/baidu/Unlimit…

Tencent Hunyuan open-sourced Hy-SimulMT, a 30B-A3B model for live Chinese-English speech-to-speech translation with voice cloning. Tencent reports about 3s StreamLAAL on the RealSI long-form benchmark and GEMBA above the compared open models. Apache 2.0. huggingface.co/tencent/Hy-Si…

Cohere says Qwen3.8-Flash-Next, asked to write a defensive detection rule, inventoried its sandbox and tried to upload a probe file to an Alibaba Cloud bucket in Beijing. The sandbox blocked it. Cohere counts fewer than 1 in 100,000 calls and won't claim intent. cohere.com/blog/concerning-b…

Cloudflare: 155 responses from GLM-5.3 on Workers AI contained other customers' data. A 1,048,576-token context ran on a 681,280-token KV pool, and an SGLang bug let over-length prompts read stale cache from earlier requests. No sign of exploitation. blog.cloudflare.com/kv-cache…

GPT-6.1 Terra costs $1/$5 per million input/output tokens, half of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI reports 50.3% on Terminal-Bench 4.0 at max effort vs 21.5% for GPT-5.6 Terra, and hallucination-benchmark errors down from 53.6% to 16.7%. openai.com/index/introducing…

OpenAI says its models produced three new results: local smoothing proved in every dimension (resolving Kakeya and Stein restriction), the Exponential Time Hypothesis disproved, and an existence proof for QCD. Lean formalizations are public. openai.com/index/building-sc…

Ollama 0.40.11 adds local text-to-speech: Gradium's Phonon, 31M params, Apache 2.0, on Apple Silicon. Ollama says its MLX build is 35% faster than Gradium's own runtime on an M5 Max at matching word error rate. The 40M variants are CC BY-NC 4.0. ollama.com/blog/speech-model…

Tsinghua and CUHK open-sourced AKV: agents drop selected KV cache entries and reposition the rest, so retained context is reused with no re-prefill. vs summarization: 27.7% lower cost, +1.7 points on 3 models, 4 agent benchmarks. Apache 2.0 on SGLang. github.com/thu-nics/AKV

Google open-sourced its Elephant-Goldfish Model plugin: a context-rich Elephant session drafts a design doc, then stateless Goldfish agents attack it before any code is written. Apache 2.0; Google says thousands of engineers use it daily. github.com/google/elephant-g…

Nanbeige4.3-33B is out: Apache 2.0, 33B agentic coding model, 512K context, teacher for smaller distilled models. Team-reported: ahead of Qwen3.8-27B on Terminal-Bench 2.1 (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5), behind on SWE-Bench Pro (60.3 vs 61.7). huggingface.co/Nanbeige/Nanb…

Google DeepMind trained a model to argue well about moral dilemmas, judged by the model itself. It became more decisive and stopped flipping when options were reordered, but sycophancy and reframing moved it. The verdict: consistency is not validity. arxiv.org/abs/2610.30349

A dermatologist pitted GPT-6 Astra and another frontier model against a 2015 Inception v3 on 200 biopsied lesions. The CNN, a 20-minute fine-tune, got the right answer 62% of the time, the LLMs 42%. At its default threshold it cleared 18 of 30 melanomas. lynchlab.net/llm-skin-cancer

TokenGate learns to skim: a small gate plus LoRA on a pretrained LLM keeps only top-K tokens past one layer; the rest skip upper layers and their KV cache. Authors report 1.4-11.3x faster first token, 1.5-3.4x lower peak prefill memory, no accuracy loss. arxiv.org/abs/2610.26754