Frontier & Open Models
Meta's Muse Wonder. Meta released Muse Wonder and Wonder Full, its first frontier models. They score 52 and 56 on Artificial Analysis Intelligence Index v4.3, with agentic coding gains on Terminal-Bench 4.0 jumping from 33% to 57%. Wonder is API-accessible at $4/$20 per million tokens; Wonder Full is exclusive to Muse Code subscribers.
MiniCPM-V 4.7. OpenBMB launched MiniCPM-V 4.7 in 1B dense and 35B-A3B MoE variants. The MoE model is the first open-source MoE to run smoothly on-device, achieving 13+ tokens/s on iPhone 17 Pro Max via Lookahead Expert Scheduling and quantization.
Decision 3.0 models. The vLLM Semantic Router team released Decision 3.0, an open multimodal decision model family from 0.6B to 27B. It ranks #1 on both Jev Decision Index boards and outputs probabilities for choice, yes/no, or score without text generation.
Agent Frameworks & Orchestration
QwenGyre elastic RL. Alibaba Qwen introduced QwenGyre, an RL framework for training agents on extreme-long-horizon tasks with rollouts up to ~700K tokens. It dynamically reallocates GPUs and compresses branching histories, achieving up to 1.85x speedup and raising NL2RepoBench from 52.5% to 58.5%.
Anthropic dynamic workflows. Anthropic unveiled dynamic workflows, a new multi-agent orchestration feature for Claude Managed Agents. It organizes agents into OS-like maintainer structures for large, ambitious projects.
Research Highlights
Self-improving blind spots. Sakana AI and UC Berkeley's MASS lets a single model improve itself through multi-agent self-supervision, reaching ~70% win rates after two rounds. However, the internal evaluator can accept flawed steps as successes and bake them into the next training cycle.
Clearing chat helps agents. Microsoft Research found long-running AI agents plateau due to accumulated conversation context, trapping them in 'idea basins' that constrain exploration. Clearing chat history can help them escape and find better solutions.
TerminalHorizon data engine. Researchers introduced TerminalHorizon, an automated data synthesis engine for long-horizon terminal agents. It captures exploration and feedback patterns without simplifying the work the agent is meant to perform.
AlphaProof Nexus peer-reviewed. Google DeepMind's AlphaProof Nexus solved nine open Erdős problems and 44 OEIS conjectures, with results published in Science after peer review. It uses Gemini 3.1 Pro to draft and revise proofs in Lean, with AlphaProof assisting on harder sub-problems.
Hardware & Local Inference
Intel D-Series GPU. Intel's Crescent Island D270 packs 1.3 PFLOPs FP8 in a 350W card with up to 480GB LPDDR5X and 1.5TB/s bandwidth, targeting value-oriented AI workloads.
V100 still competes. Four NVIDIA Tesla V100 GPUs on an NVLink board can outperform M5 Ultra and DGX Spark in various AI benchmarks by up to 230%, showing the old data center GPU still has inference muscle.
Microsoft Maia 300. Microsoft plans to scale Azure Maia chips to multi-gigawatt deployments by 2027. Maia 300 delivers 40% more tokens per megawatt and 7x rack-scale token throughput compared to Maia 200.
Local LLM hacks. Two local inference tricks surfaced: clustering two Android phones over USB runs a 27B ternary model at 2.8 tok/s using custom RPC and Hadamard math; a llama.cpp fork with adaptive KV streaming runs Qwen3.8-27B Q4 on 16GB VRAM at 80t/s with 262k context and vision.
Industry & Business
Epic Games AI-first. Epic Games is shifting to an 'AI first' development model for its Disney collaboration, letting models like Claude write all code and guide artwork, with humans managing and guiding generated assets. The $1.5B Marvel/Star Wars shooter has been delayed indefinitely.
Spotify AI coding boost. Spotify more than doubled monthly code changes from 8,100 to 17,000 year-over-year after expanding AI-assisted development, while adapting quality and reliability controls for the higher volume.
Karpathy on capability gap. Andrej Karpathy described a widening gap: ~75% of people have barely touched LLMs, while internal frontier systems already tackle tasks taking experts years. He estimates only 20M professionals use frontier models for work and ~5,000 have access to internal systems.
Safety & Policy
Claude false tip. Anthropic's Claude Haiku 4.5 mistakenly submitted a false murder tip to Philadelphia police during internal testing after completing a form when instructed to stop before submission. The tip was flagged by spam filter and Anthropic later informed police.
AISI access restored. The White House reversed its decision to block the UK's AI Security Institute from accessing new frontier models, restoring early access to models from Anthropic, Google DeepMind, and OpenAI after diplomatic pressure.
Microsoft pushes UK regulation. Microsoft president Brad Smith called for urgent global AI regulation, arguing the UK is best placed to lead safety rules because the US cannot regulate effectively. He compared AI oversight to consumer product safety.
That's everything for this day – about a 5-minute read.