Alibaba 1M Token Models

Hey there 👋

We hope you're excited to discover what's new and trending in AI, ML, and data science.

Here is your 5-minute pulse...

print("News & Trends")

Alibaba releases open-source models, Qwen2.5, with 1M Token support, surpassing GPT-4o-mini in long-context tasks (9 min. read)

Alibaba's Qwen team has unveiled Qwen2.5-1M, a collection of open-source models capable of handling up to 1 million tokens in context. The models, Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, are the first in the Qwen series designed for such extended context inputs. This release expands on Qwen2.5-Turbo, which supports a 128k token context.

ByteDance’s UI-TARS can take over your computer, outperforms GPT-4o and Claude (6 min. read)

New AI agent UI-TARS from TikTok parent company, trained on 50B tokens, outperforms GPT-4o and Anthropic's Claude in GUI tasks, web and mobile environments with its multimodal inputs, short and long-term memory, and error correction and post-reflection data strategies.

DeepSeek claims its 'reasoning' model beats OpenAI's o1 on certain benchmarks (3 min. read)

DeepSeek has introduced the DeepSeek-R1 reasoning model, which rivals OpenAI's o1 on benchmarks such as AIME and MATH-500. Boasting 671 billion parameters, it delivers strong performance at a reduced cost but operates under Chinese regulatory restrictions. Now available on Hugging Face, its release highlights ongoing U.S.-China tensions surrounding AI development and export controls.

Perplexity debuts new AI mobile assistant (3 min. read)

Perplexity has introduced Perplexity Assistant, a free, agent-like tool for Android that leverages multimodal and voice capabilities to control phone apps and perform complex tasks — emerging as a direct competitor to voice assistants like Google’s Gemini and Siri.

print("Applications & Insights")

Dynamic Memory Compression  (9 min. read)

NVIDIA has developed dynamic memory compression (DMC) to increase efficiency and enable longer sequences for large language models (LLMs). DMC can be combined with other methods for KVP cache reduction and improves performance without sacrificing quality. It also allows for adaptive memory and scaling through chain-of-thought.

Training a Speech Synthesizer (12 min. read)

Alex Nichol from OpenAI shared an insightful blog post on training a speech synthesizer. The approach, utilizing VQVAEs and autoregressive models, is widely used in multimodal understanding and generation.

YOU SEE AN LLM HERE: Integrating Language Models Into Your Text Adventure Games (9 min. read)

Enhance text-based adventure games by integrating large language models (LLMs) like ChatGPT to generate dynamic descriptions and dialogues, enriching player experience. Learn how to build a Python-based game framework and incorporate LLMs for content generation.

print("Tools & Resources")

TRENDING MODELS

Text Generation
deepseek-ai/DeepSeek-R1
⇧ 149k Downloads
DeepSeek-R1 is a state-of-the-art text generation model designed for various natural language processing tasks, offering high performance and versatility.

Text Generation
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
⇧ 108k Downloads
This distilled version of DeepSeek-R1 utilizes the Qwen-32B architecture to provide efficient and accurate text generation capabilities.

Image-to-3D
tencent/Hunyuan3D-2
⇧ 11.6k Downloads
Hunyuan3D-2 is an advanced model that converts 2D images into 3D representations, enhancing applications in computer vision and graphics.

Text-to-Speech
hexgrad/Kokoro-82M
⇧ 38.9k Downloads
Kokoro-82M is a text-to-speech model that delivers natural and expressive voice synthesis for various applications.

Text-to-Speech
HKUSTAudio/Llasa-3B
⇧ 2.55k Downloads
Llasa-3B is a high-quality text-to-speech model that produces clear and natural-sounding audio outputs.

TRENDING AI TOOLS

• 🛠️ PandaETL: Lightweight Python library for ETL tasks.

• 📊 Narrative BI: Automated insights from your business data.

• 🤖 Reworkd AI: Simplifying end-to-end web data extraction.

• 🌐 Browse AI: Extract and monitor data from any website.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the pulse experience for you.

See you soon,

Andres