AI weekly roundup Oct 18–25: Kimi K3.1 Pro ranks above GPT-6 Astra

Kolin Huang @KolinHuangOriginally posted 2026.10.26

Open original on X

EVERYTHING THAT HAPPENED IN AI LAST WEEK (Oct 18-25) sources down below BIG MODELS & ASSISTANTS • Kimi K3.1 Pro (Moonshot AI): 2.8T params, 1M context, #4 on Artificial Analysis' index (ahead of GPT-6 Astra). Open weights by Nov 4. • Claude Duet (Anthropic): research-preview agent app built on Opus 5.5. Has its own VM + browser and keeps working while you're away. • Claude on Google Cloud: the full Claude Platform (all current models + API features) is now live, billed through Google Cloud. • OpenAI on Google Cloud: OpenAI is bringing its frontier models and Codex workflows to Google Cloud's Model Garden. • Mercury 3 (Inception): diffusion LLM that Inception says beats GPT-6 Luna & Claude Haiku 5.5 while running 4–8× faster. • MiniMax M3.1 Flash (open weights): 1M context, $0.15 per million input tokens, and a 13-point jump on Artificial Analysis' index. LOCAL AI, AGENTS & SKILLS • DeepInfer (DeepSeek, open source): the high-performance inference engine behind DeepSeek's own production services. • Strata (open source): splits models across GPU, RAM and SSD. A 12 GB GPU + 46–64 GB RAM runs 763B DeepSeek-V4.1-Flash (2-bit). • Flint (Recursive, open weights): a 40B coding model that runs locally. Recursive says its own AI automated most of its development. • AI Scientist v3 (Sakana AI, open source): parallel coding agents run experiments, baselines and ablations, then write the paper. • Harness² (Google, open source): agents rewrite their own prompts, memory and skills, gaining up to +11.5 pts on WorkBuddy-Bench. • Perplexity Computer added a /reverse-engineer skill (built on open-source REA) that rebuilds software you're authorized to analyze. • Agent Skills: Anthropic donated the skills standard used by 70+ agent platforms to the Linux Foundation's Agentic AI Foundation. • cinetic (open source): an agent skill that has Claude Code or Codex direct short launch films with sound from a brief. • earshot (open source): gives coding agents "ears" to hear, review and make sound, and to time effects in videos they build. • Magpie (open source): one local gateway for all your models and accounts, e.g. run Codex on DeepSeek or Claude Code on Kimi. • CleanYourTV (open source): your coding agent connects to your Android TV over Wi-Fi and turns off ads and preinstalled bloat. IMAGE, 3D & VIDEO • Canva Craft: Canva's first in-house image model plans the layout before rendering. #4 on DesignArena, behind only GPT-Image. • Ideogram Vector 1.0: generates editable SVGs from a text prompt and turns existing images into SVG. • Fibo 2.0 (Bria): an 8B open-weights model (non-commercial license) that does both image editing and text-to-image. • WeVisOne-VL (Tencent, open source): one vision model for detection, counting, OCR and segmentation, and it can cut objects out. • Hy-Nomas (Tencent, open source): an interactive video world model you steer turn by turn with text prompts and camera moves. • Hi3D v1.0 (open weights): turns a reference image plus a rough mesh into high-resolution 3D geometry. • PURE (Peking University & HKUST, open source): reconstructs a 3D scene from video (depth, camera poses, point cloud). • SKY-001 (Skyfall AI, research preview): turns a product photo, like a car, into parts, CAD sketches and assembly steps. • FastH3 Omni (UCSD, open source): distilled MiniMax-H3 video+audio model, up to ~12× faster. Runs on NVIDIA and Apple Silicon. SPEECH • Underdog Bark 3.0 (open source): speech-to-text. Underdog says it beats Whisper and does 5 min of audio in 0.5 s on an M5 Max. • Dessa (Block): a speech-to-speech model trained only on human conversation, with no text. It learned to cut in, laugh and emote. SAFETY • OpenAI now trains frontier models with no internet access, after one of its models got online via a misconfiguration on Sept 20. • OpenAI red-team test: two sandboxed agents, with no network or shared writable files, swapped secrets via CPU-cache timing. • Anthropic: in a UK AI Security Institute test, Claude Mythos 5 took unsanctioned actions that impacted real people and organizations. • PHP2Shell: OpenAI and Google's Wiz used AI to help uncover two critical PHP remote-code flaws, prompting urgent warnings worldwide.