LLM post-training - Page 6
Stories from LLM post-training come first, newest first, then stories from connected topics (Kimi K3, Qwen3.8-27B, reinforcement learning, AI Models, coding agents, open weights), each labelled with where it comes from. This view is not indexed by search engines.
Latest
Microsoft Open-Sources Mage-VL-Cowork, a 4B Agent That Clicks or Codes
Microsoft's open Mage-VL-Cowork is a 4B computer-use agent that switches between clicking and code. Microsoft reports 66% on OSWorld-Verified.
Google Is Shutting Jules; Sessions End 11 November, Export Opens Now
Jules shutdown: Google's coding agent goes read-only after 11 November and closes on 31 December. How to export sessions, memories and settings now.
Factory Says Droid 1.0 Beats Codex and Claude Code in Its Own Tests
Factory's Droid 1.0 coding-agent harness makes scripted tool calls the default. Its own Terminal-Bench 4 runs put it ahead of Codex and Claude Code.
Pika Opens H3 Flex Video Weights, but Licence Excludes UK and US
Pika's H3 Flex video model, built on MiniMax H3, has downloadable weights, but its licence excludes the US, UK, EU and South Korea.
MLCommons Ranks 27 AI Systems on Safety; Open Models Span 4th to 98th
MLCommons' AILuminate Safety 1.2.1 puts 22 open-weight models between the 4th and 98th percentile. Jailbreak 1.1 scores follow after developers are told.
OpenAI Rebuilds Its Model Spec Around Five Core Values for GPT
OpenAI has reorganised its Model Spec around five values, adding rules on human oversight and context and dropping most of its style guidance.
GitLab Orbit Goes GA, Claiming Up to 45x Fewer Agent Retries
GitLab Orbit, a context graph for AI coding agents, is now GA. GitLab says its tests showed up to 45x fewer retries and 4.5x fewer tokens.
Mistral Opens Large 4 Weights: a 1-Trillion-Parameter Model
Mistral has released the weights of Large 4, a 1-trillion-parameter multimodal model, and warns its cyber skills carry a potential for real attack uplift.
Kakao Open-Sources Two Kanana-2 Embedding Models for Korean Search
Kakao has released 3bn and 1.3bn-parameter Kanana-2 embedding models on Hugging Face, topping its Korean retrieval tests while trailing Qwen3 in English.
SenseTime Releases Code2Video, an Open Model That Makes Videos in Code
SenseTime's open SenseNova 6.8 Code2Video, a 27B model built on Qwen3.8-27B, writes and renders code to make short videos. It has no benchmarks yet.
Latest on TSN
Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.
