MC-Sparse: Training-Free 1.8x Sparse Attention Speedup
A training-free sparse attention framework from Tencent Hunyuan hits 1.80x denoising speedup on MiniMax-H3-Base and 2.32x on 3D asset generation.
Learn2Play: Opus 5.5 beats human-top average, not peak
Learn2Play Bench: 20 novel text games test learning by interaction. Opus 5.5 beats human-top on average (61.0 vs 53.5) but peaks 4.2 below.
LoRA over GGUF: Train a 125B MoE in 40 GiB VRAM
An open-source recipe runs LoRA over GGUF bases, bypassing bitsandbytes' MoE gap: Qwen3.8-Flash-Next 125B trains in 40 GiB VRAM at 200 tok/s locally.
Westlake-Ant SparseDecoding: Decoding-Aware Pruning Speeds LLMs
SparseDecoding prunes LLMs on self-generated tokens: Qwen3-14B writing 1.85→4.18, A100 decoding 1.48x faster via a bitmask SpMV kernel.
AgentGarten: Worlds as Code, Agents Learn in 4 Rounds
MirroS-Lab's AgentGarten couples simulators with a shared neural renderer; agents learn in 4 rounds vs millions for RL. Renderer code is open.
Microsoft Decision-1: 35x faster than Sol, output free
Microsoft's Decision-1: calibrated decision scoring for agent routing and guardrails. Qwen3.5-9B, topped 36 benchmarks, 35x faster vs Sol at $0.042/M.
Xiaomi published the full RL training ledger for MiMo-V2.6
Xiaomi's MiMo-V2.6 report opens its RL ledger: $2.6M for Pro, 12.7% of compute on graders, DeepSWE 58.4 to 72.6, plus a 7k-task open RL stack.
Claude Code as Forensic Tool: Korean Bank Breach
ARTEX + Claude Code + DeepSeek + GLM-5.3 + Grok 4.6 hit Korean banks; Claude Code logs leaked the attacker. Agentic logs = new forensic channel.
Gemini app tiers reshuffle: free capped at Flash-Lite
Google: Oct 9 free Gemini users limited to Flash-Lite; AI Plus loses Pro; AI Pro gains Deep Think, AI Ultra only before. Effort levels added.
xAI's Grok Imagine Video 1.5 Lite undercuts Veo 3.1 by 65%
xAI released Grok Imagine Video 1.5 Lite, distilled from 1.5. USD 0.14/sec at 1080p, 65% cheaper than Veo 3.1 with audio. AA-Video-T2V rank No. 17.
Google's Project Suncatcher prototype reaches orbit
Project Suncatcher prototype reached orbit with 4 Trillium TPUs running Gemma, capped at 15 minutes per session by vacuum cooling. Built with Planet.
TokenRouter: token-level LLM routing gets a serving engine
Tsinghua's TokenRouter: per-model subservers + delayed batching serve token-level LLM routing at 2.01-64.15x throughput. NeurIPS 2026, open source.
Mellum2.1: RL-Retrained 12B Model Now Fixes Its Own Bugs
JetBrains Mellum2.1: 12B MoE retrained with a summer of RL — now explores repos, edits files, checks its fixes; ~2x Qwen3.5-9B throughput, 1.6x via MTP.
Xiaomi Sells Inference Speed: Same Weights, 10x Price
Xiaomi's MiMo-V2.6-Pro-UltraSpeed serves the same 1T checkpoint ~10x faster at 10x the price — worthless for batch jobs, decisive for agent loops.
One Verb Flips Tool Calls: LLM Decision Vector Found
NeurIPS 2026: swapping one verb flips tool-call decisions in agentic LLMs; the choice traces to a causal vector, replicated across seven models.
EngramEdit: surgical knowledge edits via conditional memory
arXiv 2610.10533: edit DeepSeek Engram conditional memory, no retraining. ~100% edit success, ~3x baseline on CoT multi-hop.
Korean banks hit by AI-agent attack; Claude leaks operator
On October 7 CrowdStrike disclosed an AI-agent intrusion on South Korean banks. ARTEX chained with Claude Code; a résumé request exposed the operator.
Anthropic's Cyber Mission Puts Claude in OT and OSS Defense
Anthropic's Cyber Mission: Claude in OT defense (11 partners) and OSS Scanner, free open-source scans via Claude, 90% true-positive target.
Recurrent Looped Transformer: Depth That Grows With Sequence
Princeton's RLT splits layers between a parallel encoder and a recurrent decoder, hitting 100% parity at 256 bits where a Transformer stays at 50%.
STEPQuant: 6-bit Recurrent States Match FP32, Memory Down 68.7%
ZJU's STEPQuant quantizes Delta-rule recurrent states; 6-bit matches FP32, 4-bit beats INT8 on Qwen3.8-27B, SGLang serving memory down 68.7%.