AI Daily · Nov 9, 2026: Meituan open-sources LongCat-2.5 weights under MIT, 1.6T-parameter natively multimodal model
Meituan released LongCat-2.5 weights on Hugging Face under MIT (1.6T total, ~48B active), while OpenCode offers it free for two weeks; Mistral Cloud added Apps and Managed Agents previews.
Showing the latest briefing (Nov 9, 2026). Every day is generated the next morning.
Pick a date
-
Meituan releases LongCat-2.5 weights under MIT license
LongCat-2.5 is a natively multimodal agentic model with 1.6T total parameters and about 48B active, released under MIT with weights on Hugging Face.
-
LongCat 2.5 available free on OpenCode for two weeks
OpenCode is offering LongCat 2.5 free for two weeks, with 1M context, multimodal input, and zero data retention.
-
Mistral Cloud adds Apps and Managed Agents previews, Workflows GA
Mistral Cloud now offers Apps and Managed Agents in public preview and Workflows in general availability, with private cloud and on-premises deployments arriving at the end of November.
-
Command Code adds free stealth model Crouton Alpha with 1M context
Crouton Alpha, a free multimodal reasoning stealth model with 1M token context, is available in Command Code via --model stealth/crouton-alpha; the same recap also lists DeepSeek V4.1 Flash 1017 Ultrafast at 1K+ TPS and GLM 5.3 Flash at 300-500 TPS.
-
Qoder offers stealth Qoduck model free for limited time
Qoder's stealth model Qoduck is free for a limited time; community builds include 68 walkable rooms in a voxel White House, a floating palace in one HTML file, and a QR-scanner Chrome extension shipped same day. Qoder Sites now supports 3D scenes, exploded product views and in-page games via the /sites command.
-
Liquid AI harness adapters trained on LFM2.5-24B and LFM3 boost tool use
Harness adapter training finished for Liquid AI's LFM2.5-24B and LFM3 plus thewebAI's Twil-8B, yielding up to 43% better tool efficiency and 15% accuracy gain on a held-out data-analysis set for LFM3, and up to 37% tool-call efficiency on Twil-8B after over 1,900 RL steps on Hugging Face jobs.
-
OpenCode 2 climbs from eighth to first on Hugging Face harness leaderboard
OpenCode 2 replaces version 1.19 on the Hugging Face Harness Leaderboard, moving from 8th of 9 harnesses to tied for first across all three models and solving 65.7% of tasks at high effort.
-
RicursiveAI releases Ripple-65B Preview dense open-weight LLM
Ripple-65B Preview is a 65B dense model pretrained from scratch that claims state-of-the-art agentic coding results among U.S. open-weight models, beating Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B.
-
LiteLLM Gateway dashboard shows per-deployment pricing across billing modes
The Model Hub in the LiteLLM Gateway Dashboard now displays pricing for each model deployment, covering cached, long prompt, flex, priority and batch modes in one place.
-
Mistral says Studio customers can run models in own environment
Mistral's Applied AI team builds customer solutions on Studio, including penetration testing and KYC onboarding, and customers can run both models and platform in their own environment to keep sensitive data in-house.