Easy AI Daily | November 27, 2025

Key points

• Anthropic published persistent agent best practices and MCP released the SEP-1686 "tasks" protocol for background long-running tasks
• Claude Opus 4.5 tops LisanBench and Code Arena WebDev, though the non-Thinking variant shows regression
• Alibaba open-sources Z-Image-Turbo, a 6B text-to-image model approaching Seedream 4.0 quality
• LatentMAS reduces multi-agent communication tokens by 70-84% with 4-4.3x speedup
• Hugging Face data shows Chinese models now account for 17.1% of downloads, led by DeepSeek and Qwen

Agent & Tool Ecosystem

Anthropic releases persistent agent framework and MCP tasks protocol update

Anthropic proposed persistent agent practices (state checkpoints, structured artifacts); MCP released SEP-1686 "tasks" supporting background long-running tasks; LangChain clarified the framework-runtime-harness stack, with LangGraph classified as a runtime.
• Anthropic blog summary | MCP tasks announcement | LangChain stack explanation

Booking.com deploys production agents for customer messaging

Booking.com built agents with LangGraph and Kubernetes, handling tens of thousands of messages daily with 70% satisfaction improvement, using GPT-4 Mini and Weaviate for semantic search.
• Technical deep dive

Perplexity launches Memory and virtual try-on

Perplexity added user-level Memory (view/delete/disable supported) and a virtual try-on shopping feature.
• Memory announcement | Virtual try-on

Model Updates & Performance

Claude Opus 4.5 shines in benchmarks

Opus 4.5 Thinking ranks first on LisanBench and tops Code Arena WebDev; the non-Thinking version regressed, with community reports of Python tool misuse; Claude.ai now auto-compresses context.
• LisanBench results | Code Arena leaderboard | Context compression update

Alibaba open-sources Z-Image-Turbo text-to-image model

Alibaba released Z-Image-Turbo (6B parameters, based on a Qwen3 4B text encoder), free on ModelScope and integrated into Hugging Face Diffusers, performing close to Seedream 4.0.
• ModelScope | Reddit discussion

FLUX.2 series released

FLUX.2 pro/flex models joined LMArena; users report improved visual quality with the "plastic look" eliminated, competitive against NB Pro.
• LMArena announcement | Comparison

EGGROLL improves evolution strategies efficiency

EGGROLL uses low-rank perturbations to accelerate evolution strategies, supports 100k+ populations, and stably pre-trains recurrent LMs for large discrete systems.
• Technical overview

dnet solves Apple Silicon memory limits

Dria's dnet uses distributed inference, disk streaming, and UMA scheduling to run over-memory models on Apple Silicon clusters, addressing OOM issues.
• Announcement

Inference & Efficiency

LatentMAS cuts multi-agent communication tokens

LatentMAS replaces text communication with latent vectors, reducing tokens by 70-84% and improving speed 4-4.3x without accuracy loss.
• Paper | Summary

Reasoning trace distillation lowers costs

Training a 12B model on gpt-oss traces reduces token usage 4x and cuts costs by avoiding repeated inference.
• Summary | Demo

Multimodal & Generative Models

PixelDiT optimizes pixel diffusion

PixelDiT uses dual Transformers (patch-level and pixel-level), achieving ImageNet 256x256 FID 1.61 and GenEval 0.74.
• Paper

Apple releases STARFlow-V video generation model

STARFlow-V uses normalizing flows, supports T2V/I2V/V2V with causal prediction, and flow-score matching improves consistency.
• Paper

Nano Banana 2 improves structured images

Nano Banana 2 shows improved StructBench performance; community shared prompt resources.
• Analysis | Resources

Open-Source Ecosystem & Evaluation

HF download data shows Chinese model growth

Chinese models account for 17.1% of downloads, surpassing the US, led by DeepSeek and Qwen; multimodal models are popular.
• Overview | Thread

METR becomes a trusted evaluator

METR is regarded by practitioners as the most trusted external evaluator for model performance verification.
• Comment

AI Security Institute publishes Anthropic case study

Evaluated whether Opus 4.5 would sabotage AI safety research; results were positive with caveats.
• Thread

Zhihu optimizes multimodal recommendations with Qwen2.5-VL

A pipeline driven by Qwen2.5-VL-72B/3B with LoRA fine-tuning improves MMEB-eval-zh by 7.4% over embedding baselines.
• Write-up

New benchmarks released

MultiPathQA (pathology navigation), MTBBench (oncology decision-making), and WER is Unaware (clinical ASR) launched.
• Pathology | MTBBench | WER

Reddit Highlights

• Z-Image-Turbo discussion: users note performance near Seedream 4.0; the 6B size suits local deployment. Reddit
• Opus 4.5 converts ZBar to Swift 6: solved a long-standing bug other models failed at. Reddit
• Opus 4.5 benchmark chart controversy: 80.9% SWE-bench accuracy leads, but chart design was criticized. Reddit
• AI progress chart debate: Thomas Pueyo's "fun toy to AGI" chart questioned for professionalism. Reddit
• AI memes: Ilya Sutskever scaling quotes, Grok 4.1's unhinged replies, Gemini 3 satire. Singularity | ChatGPT

Discord Community Discussions

• LMArena: Flux 2 vs NB Pro comparisons; SynthID prevents nerfing. Announcement
• Unsloth AI: ERNIE developer challenge support (Devpost), ES HyperScale CPU training, Qwen3 fine-tuning issues
• GPU MODE: Triton kernels, NVFP4_GEMV leaderboard, multi-node inference (paper)
• OpenRouter: Opus overload, DeepSeek R1 delisting, model fallback logic bugs (docs)
• Eleuther: multi-stage LLM hallucinations, SGD shuffling controversy, Emergent Misalignment replication (paper)
• Latent Space: Claude Code Plan Mode upgrade, Jeff Dean's 15-year ML retrospective
• DSPy: dspy-cli open-sourced with FastAPI and MCP support (repo)
• LM Studio: API endpoint fixes, image captioning workarounds (docs)
• tinygrad: TinyJit kernel replay and random function implementation (tutorial)

*Source: Easy AI education project*