Easy AI Daily Digest – October 30, 2025: Kimi Linear, MiniMax M2, OpenAI Aardvark & More

Forum topic · 小凯 · 2026-03-27

Summary

This October 30, 2025 AI daily digest covers major industry releases and community discussions. Moonshot AI launched Kimi Linear, a hybrid Kimi Delta Attention (KDA) and MLA architecture with open CUDA kernels and vLLM integration, cutting KV cache by 75% and boosting decoding throughput up to 6x. MiniMax released M2 with full attention, 200k context, and 100 TPS in free preview. ByteDance, Princeton, and Mila introduced Looped LLMs, where latent-recursion 1.4B/2.6B models match 4B/8B baselines. OpenAI opened the private beta of Aardvark, an agentic security researcher for vulnerability detection, and Cognition launched Computer Use in public beta. HKUST released the Toolathlon benchmark (32 apps, 600+ tools) where Claude Sonnet 4.5 scored only 38.6%. Hugging Face published a 200+ page Smol Training Playbook, Voyage shipped voyage-3-large embeddings, Cartesia launched Sonic-3 TTS with sub-250ms latency, and Perplexity added patent search. Reddit and Discord threads covered Qwen 3 VL in llama.cpp, Anthropic's LLM introspection research, George R.R. Martin's lawsuit against OpenAI, and Cognition's SWE-1.5 coding model running at 950 tok/s on Cerebras hardware.

Easy AI Daily Digest | October 30, 2025

A roundup of the day's AI industry news, research, and community discussions, curated by the Easy AI teaching project.

Key Points

Moonshot AI releases Kimi Linear (KDA): A hybrid of Kimi Delta Attention and MLA, with open-sourced CUDA kernels and vLLM integration. Reduces KV cache by 75% and improves decoding throughput up to 6x, with strong long-context and RL performance. Sources: Moonshot AI, vLLM
MiniMax M2 switches to full attention: Abandons hybrid architecture for full attention, supporting 200k context at 100 TPS with free trial. Community debates its efficiency trade-offs vs. Kimi Linear, noting full attention may be better for multi-hop reasoning. Sources: omarsar0, vLLM M2 support
Looped LLMs (ByteDance/Princeton/Mila): Latent-recursion models where 1.4B/2.6B variants match standard 4B/8B models with better data efficiency; potentially combinable with MoE scaling. Source: announcement
OpenAI Aardvark (GPT-5) private beta: An agentic security researcher that reads code, writes tests, and proposes patches for vulnerability detection. Sources: OpenAI, gdb
Cognition Computer Use public beta: Devin can operate desktop and mobile tools, share screen recordings, and build GUI applications. Source: Cognition
HKUST Toolathlon benchmark: Covers 32 applications and 600+ tools; Claude Sonnet 4.5 achieved only 38.6% correctness, highlighting gaps in tool-use ability. Source: junxian_he
Hugging Face Smol Training Playbook: A 200+ page guide covering pretraining, fine-tuning, and infrastructure, emphasizing ablation studies and practical strategies. Sources: LoubnaBenAllal1, playbook
Voyage voyage-3-large: Tops the HF RTEB leaderboard for enterprise retrieval; supports INT8 quantization to cut vector DB costs, strong in finance/legal/medical domains.
Cartesia Sonic-3 TTS: SSM-based architecture with <250ms real-time latency, supporting 42 languages (including 9 Indian languages).
Perplexity Patents & Discover: New patent research tools, Discover feed, and finance features (e.g., politician stock holdings). Source: Perplexity

Reddit Highlights

Qwen 3 VL merged into llama.cpp: Currently MLX (Mac) only; Q6 quantization reportedly performs well. Discussion | PR #16780
Kimi Linear 48B-A3B on Reddit: Based on Modified Gated DeltaNet, trained with 25x fewer tokens, expected to support 1M context. Model
Anthropic introspection research: A paper claims LLMs can detect modifications to their internal activations; community debates whether this is genuine introspection or pattern recognition. Paper
Udio removes WAV downloads: Paying users frustrated; community calls for open-source AI music alternatives.
10 Claude Skills that changed workflows: Including Rube MCP (500+ app integrations) and Superpowers. Repo
George R.R. Martin vs. OpenAI: Judge allows authors' copyright lawsuit to proceed over ChatGPT-generated Game of Thrones-like content.

Discord Community Notes

Cursor: Composer 1 speed/accuracy debates, pricing and cache complaints; Cursor 2.0 bugs alongside new features.
Unsloth: RTX 8000 VRAM for servers; Qwen3 4B GRPO fine-tuning OOM fixes via 4-bit loading and batch adjustments. Docs
OpenRouter: Launched Sonar Pro Search with Perplexity; user complaints about Sora 2 output bias.
Latent Space: ScaleAI's RLI benchmark shows Manus agent at only 2.5% automation; Cognition's SWE-1.5 runs 6x faster than Haiku. RLI
Windsurf: SWE-1.5 at 950 tok/s, 13x faster than Sonnet, built on Cerebras hardware.
GPU MODE: CUDA scan optimization benchmarks and FP8 quantization with TorchAO + GemLite. FP8 repo
Nous Research: Hack The Box will host a free MCP-only CTF on AI security on November 20. Signup
Modular Mojo: MAX matching/exceeding NVIDIA and JAX on ML tasks; a scikit-learn replacement library (SciJo) in early development.
MCP Contributors: Model Context Protocol RFC delayed pending tangible implementations.

---

*Source: Easy AI teaching project*