Easy AI Daily News | November 5, 2025: Kimi-K2 in vLLM, Perplexity MoE Kernels, Anthropic Tool-Calling Guide

Model Integration & Deployment

• Kimi-K2 reasoning model merged into vLLM, with SGLang support planned. Its MoE configuration has roughly 1.2T total parameters / 30B active parameters, similar to other recent large sparse models.
• Perplexity released custom MoE kernels and a research paper supporting large MoE deployments (e.g., Kimi K2) on AWS EFA; vLLM hinted it will integrate the fast communication kernels.
• vLLM v1 adds first-class hybrid model support (dense + sparse experts) via IBM's vLLM team, covering Qwen3-Next, Nemotron Nano 2, Granite 4.0, and others.
• Unverified Kimi-K2 benchmark claims: a reported 77% on GPQA Diamond (vs. GPT-4.5's 71.4%) awaits broader evaluation.

Agent Systems & Tools

• Anthropic published a tool-calling optimization guide: MCP servers as code APIs, progressive tool discovery, and in-environment data processing cut context from 150k to 2k tokens.
• Graphiti MCP server connects Claude Desktop and Cursor for fully local, cross-app temporal knowledge-graph memory sharing.
• VS Code introduced an "Agent sessions" view to manage Copilot and external agents (e.g., Codex) from one place.
• Cursor improved large-codebase performance with trained code-retrieval embeddings, reporting semantic search outperforms grep.
• Agent eval updates: CodeClash runs multi-round code duels; LMArena launched "Arena Expert," a profession-labeled leaderboard based on real user traffic.

Multimodal & Video Generation

• ByteDance released BindWeave for subject-consistent image-to-video generation via cross-modal integration; model card available on Hugging Face.
• MotionStream achieved ~29 FPS with 0.4s latency on a single H100, enabling interactive motion-controlled real-time video generation.
• Google Veo 3.1 "Camera Adjustment" allows adjusting angles/motion of generated videos; a Qwen Image Edit Multiple Angles LoRA adds camera pose control.
• New benchmarks/tools: ViDoRe v3 (real-world multimodal RAG eval), VCode (visual-to-SVG code), MIRA (visual chain-of-thought testing).

Research & Training

• OpenAI released IndQA, a benchmark evaluating AI understanding of Indian languages and everyday cultural context.
• Learning rate transfer under μP theory formally proven, advancing the theoretical foundations of model scaling.
• Anthropic observed unreliable introspection in LLMs via "concept injection," detecting internal thoughts vs. inputs and intended vs. unintended behavior.
• Edison Scientific's Kosmos ran 200 agent rollouts, executed 42k lines of code, read 1.5k papers, and reported 7 externally validated discoveries (metabolomics, materials, etc.).
• NVFP4 quantization progress: custom Cutlass kernels beat cuBLAS; global/local scaling and calibration workflows; Wan 2.2 approaches bf16 quality under NVFP4.

Ecosystem & Platforms

• OpenAI COO says 1M+ businesses use its products; "OpenAI for Science" launched, positioning GPT-5 as a domain research collaborator.
• Perplexity becomes Snapchat's default AI starting January 2026.
• Gemini expands into Google products: Deep Research can pull Workspace data for reports; Gemini enters Google Maps for hands-free route queries.
• Other updates: OpenHands Cloud free base tier; openenv for push/pull RL environments; Voiceflow KB metadata routing; Dify integrates Qdrant for RAG; Nebius Token Factory; LlamaBarn v0.10.0 beta.

Community Discussions (Reddit)

• Qwen model usability complaints (sycophancy, quantization specs, prompting for skepticism)
• Local AI hardware setups: PCIe bifurcation, GPU choices (A6000, A40, 3090), cost/performance tradeoffs
• Anticipation for GLM 4.6 AIR and comparisons with GLM 4.5 AIR
• XPENG's new humanoid robot design (chest cooling, human-like appearance)
• Gemini 3 rumors (alleged 1.2T parameters) and Apple's Siri reportedly powered by Gemini
• AI art: award-winning AI short film (Indian AI film festival), AI game playthroughs, Llama3 existence-reflection projects

Discord Highlights

• LM Studio 0.3.31: faster VLM OCR, Flash Attention default on CUDA GPUs, MiniMax-M2 tool calling, new lms runtime CLI
• Perplexity model confusion: users report Haiku or Gemini 2 Flash responses when selecting Claude Sonnet 4.5 or Gemini 2.5 Pro
• Hugging Face acquired Sentence Transformers and released huggingface_hub v1.0
• Sora app on Android (Canada, Japan, etc.); IndQA benchmark launched
• tinybox pro v2 released: 8x 5090 GPU 5U rackable workstation, $50,000, 4–12 week shipping
• Windsurf released Codemaps built on SWE-1.5 and Sonnet 4.5 for improved code understanding
• Unsloth AI released DeepSeek-OCR fine-tuning notebooks (users reported error rates over 100%)
• Stability AI won its lawsuit against Getty Images; RWKV progress discussed (HRM/TRM merge)
• tinygrad, GPU MODE, Eleuther, DSPy, MCP Contributors: discussions on CUDA memory-bound matmuls, AMD/NVIDIA kernel competition entries, concept detection/steering systems, DSPy pause/resume optimization, and IETF 124 OAuth topics for AI crawlers

*Source: Easy AI education project.*