HomeTagsLLM post-training

LLM post-training - Page 4

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

Anthropic Launches Claude Fable 5.2 and Mythos 5.2

Anthropic launched Claude Fable 5.2 for all users and Mythos 5.2 via verification. Benchmarks are Anthropic’s; Opus 5.5 stays the daily driver.

Meta Releases Muse Ember, Its Smallest Agent Model

Meta released Muse Ember, its smallest Muse model for agent work, at $0.10/$0.50 per million tokens. Benchmarks are Meta’s; Haiku leads FrontierCode.

LG AI Research Releases Free API Access to K-EXAONE 2.5

LG AI Research has opened a free API for K-EXAONE 2.5. Internal agentic scores are company-reported and trail GLM-5.3 on two of four tests.

MOSS-Live Links Full-Duplex Voice, Vision and Background Tasks

MOSS-Live combines full-duplex speech, live vision and background tools. OpenMOSS’s own tests show strengths in interaction, but mixed task results.

OpenRouter Launches Auto Code and Rebuilds Its Auto Router

OpenRouter Auto Code selects models for coding agents. The company reports lower cost per task in five benchmarks; model usage is still charged.

Carta Evergreen Opens Multiplayer AI Coding for Teams

Carta Evergreen: multiplayer AI coding with agents; Carta cites ~20k PRs/week and lower reverts. Usage fees: $0.15/M tokens plus provider price.

Cognition Says Devin Review 2.0 Cuts PR Review Cost 82%

Devin Review 2.0: Cognition says PR review cost down 82% since August; typical $0.25–$0.75. Company metrics, not independent benchmarks.

OpenAI Launches GPT-6 Cyber for Defensive Research

GPT-6 Cyber: OpenAI says its cyber model scores 48.5% on ExploitGym and helped produce 948 Linux LPE exploits in a controlled research pipeline.

Google Makes Gemini Omni 1.2 Flash Generally Available

Gemini Omni 1.2 Flash is generally available: native 15s video, $0.07/s at 720p (30% less than 1.1), multi-turn editing, SynthID.

Baidu Releases ERNIE-5.5-Flash-Base as Open Weights

Baidu releases ERNIE-5.5-Flash-Base open weights: 280B MoE with 14B active, 33.5T tokens, licence excluding the EU, UK and South Korea.

Latest on TSN

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.