Mistral AI released Small 3.1, a SOTA multilingual and multimodal LLM —24B (can run on a laptop)
—128k token context window —Outperforms Gemma 3 and GPT-4o Mini on most benchmarks
—Inference speed of 150 tokens/sec
—Open-source under Apache 2.0 license
@rowancheung
-
Mistral AI Releases Small 3.1 Multilingual Multimodal LLM
By
–
-
Roblox Launches Cube 3D: AI-Powered Text-to-3D Generation Tool
By
–
Roblox announced Cube 3D, an open-source AI for generating 3D objects and scenes from text prompts
— Rowan Cheung (@rowancheung) 18 mars 2025
Trained on native 3D data, Cube uses 3D tokenization to create functional 3D outputs in seconds
Support for image inputs is also on the way!pic.twitter.com/dxl5jNumPvRoblox announced Cube 3D, an open-source AI for generating 3D objects and scenes from text prompts Trained on native 3D data, Cube uses 3D tokenization to create functional 3D outputs in seconds Support for image inputs is also on the way!
-
Roblox Simplifies 3D Content Creation With AI Text Commands
By
–
TODAY'S AI NEWS: Roblox just made 3D content creation as simple as a text command! Plus, more news from Mistral, xAI, Zoom, Tencent, Perplexity, Baidu, and more. Here's everything you need to know:
-
Manus Partners with Alibaba Qwen for Chinese Autonomous Agent
By
–
The team behind Manus partnered with Alibaba's Qwen to develop a Chinese version of its autonomous agent
— Rowan Cheung (@rowancheung) 12 mars 2025
The collaboration will integrate Manus with Qwen's open-source models and computing infra
This comes just after Manus went viral last week!pic.twitter.com/encoTIamO3The team behind Manus partnered with Alibaba's Qwen to develop a Chinese version of its autonomous agent The collaboration will integrate Manus with Qwen's open-source models and computing infra This comes just after Manus went viral last week!
-
Luma Labs Breakthrough: Inductive Moment Matching Improves Image Generation
By
–
Luma Labs introduced a new pre-training technique called Inductive Moment Matching
— Rowan Cheung (@rowancheung) 12 mars 2025
It produces superior image generation quality 10x more efficiently than current approaches
Luma says the approach breaks the algorithmic ceiling of diffusion models!pic.twitter.com/h0x8kG18i4Luma Labs introduced a new pre-training technique called Inductive Moment Matching It produces superior image generation quality 10x more efficiently than current approaches Luma says the approach breaks the algorithmic ceiling of diffusion models!
-
Harvey Launches AI Legal Workflow Agents Matching Human Lawyer Performance
By
–
Harvey released Workflows AI agents for legal tasks, with reasoning, planning, and adapting capabilities
— Rowan Cheung (@rowancheung) 12 mars 2025
In blind reviews, lawyer evaluators rated legal work produced by workflow agents as equal to or better than that of human lawyerspic.twitter.com/8DTXxDzVzGHarvey released Workflows AI agents for legal tasks, with reasoning, planning, and adapting capabilities In blind reviews, lawyer evaluators rated legal work produced by workflow agents as equal to or better than that of human lawyers
-
Cartesia Sonic 2.0: $64M Funding, Multilingual AI Voice Model
By
–
Cartesia announced $64M in funding and a new Sonic 2.0 state space model
— Rowan Cheung (@rowancheung) 12 mars 2025
The model generates realistic AI voice across 16 languages and also includes an ultra-fast 'Turbo' mode
In blind tests, 1.5x as many people preferred Sonic over other providerspic.twitter.com/nciAMFRPt0Cartesia announced $64M in funding and a new Sonic 2.0 state space model The model generates realistic AI voice across 16 languages and also includes an ultra-fast 'Turbo' mode In blind tests, 1.5x as many people preferred Sonic over other providers
-
Reka Open-Sources Flash 3: 21B Reasoning Model for On-Device AI
By
–
Reka open-sourced Flash 3, a 21B parameter reasoning model
— Rowan Cheung (@rowancheung) 12 mars 2025
—Matches o1-mini on benchmarks
—32k context length
—Ideal for on-device deployment
—Finetuned on synthetic and public datasets
—Powers Nexus, Reka's platform for creating and managing AI agentspic.twitter.com/FTzKc8NtM8Reka open-sourced Flash 3, a 21B parameter reasoning model —Matches o1-mini on benchmarks
—32k context length
—Ideal for on-device deployment
—Finetuned on synthetic and public datasets
—Powers Nexus, Reka's platform for creating and managing AI agents -
Meta Develops In-House AI Chip to Reduce Training Costs
By
–
Meta is testing a new, in-house chip to cut costs on AI training
— Rowan Cheung (@rowancheung) 12 mars 2025
Manufactured by TSMC, the chip is part of the company's MTIA series and is likely to be deployed in 2026
It will help Meta cut reliance on Nvidia's pricey GPUs for training large modelspic.twitter.com/Blpj3jusDjMeta is testing a new, in-house chip to cut costs on AI training Manufactured by TSMC, the chip is part of the company's MTIA series and is likely to be deployed in 2026 It will help Meta cut reliance on Nvidia's pricey GPUs for training large models
-
OpenAI Launches DIY Custom AI Agent Building Tools and SDK
By
–
OpenAI released new DIY tools for building custom AI agents, including:
— Rowan Cheung (@rowancheung) 12 mars 2025
A CUA model-powered Responses API with tools for web browsing, computer use, and file management
An open-source Agents SDK to orchestrate single and multi-agent systemspic.twitter.com/70QcVRL3EFOpenAI released new DIY tools for building custom AI agents, including: A CUA model-powered Responses API with tools for web browsing, computer use, and file management An open-source Agents SDK to orchestrate single and multi-agent systems