Alibaba/Qwen

Latest AI news, models and releases from Alibaba/Qwen. ['CodeQwen1.5', 'Cotype Light 3', 'Cotype Pro 3', 'Fun-Realtime-TTS', 'Hanguang', 'Hanguang 800', 'HappyOyster 1.0', 'ICN Switch 1.0', 'not specified', 'Panjiu', 'Qianwen', 'Qoder Security', 'QVQ-72B-Preview', 'QVQ-Max', 'Qwen', 'Qwen 2', 'Qwen2.5', 'Qwen2.5-0.5B', 'Qwen2.5-14B', 'Qwen2.5-14B-Instruct-1M', 'Qwen2.5-1.5B', 'Qwen2.5-32B', 'Qwen2.5-3B', 'Qwen-2.5 72B', 'Qwen2.5-72B', 'Qwen2.5-7B', 'Qwen2.5-7B-Instruct-1M', 'Qwen2.5-Coder', 'Qwen2.5-Coder-0.5B', 'Qwen2.5-Coder-0.5B-Instruct', 'Qwen2.5-Coder-14B', 'Qwen2.5-Coder-1.5B', 'Qwen2.5-Coder-1.5B-Instruct', 'Qwen2.5-Coder-32B', 'Qwen2.5-Coder-32B-Instruct', 'Qwen2.5-Coder-3B', 'Qwen2.5-Coder-3B-Instruct', 'Qwen2.5-Coder-7B', 'Qwen2.5-Coder-7B-Instruct', 'Qwen2.5-Coder-Instruct']

Official site ↗

Research 🇷🇺

Token Cramming: Ideal Text Compression Does Not Preserve Meaning – AIRI Study

Researchers at AIRI demonstrated that the token cramming method, which compresses up to 1568 tokens into a single embedding, does not guarantee meaning preservation: under greedy decoding, accuracy drops from 99.96% to 40%. A new method, progressive cramming, was proposed to achieve 100% reconstruction, but it was found that even a perfectly compressed embedding impairs the model's reasoning ability, and the cause lies in the early layers of the transformer.

Context Engineering: What It Is, How to Work with LLM Agent Context, and Why You Get Paid for It

Context engineering is a set of techniques for collecting and maintaining optimal context when working with LLMs. It differs from prompt engineering (which only handles instructions) and complements RAG (knowledge delivery) and MCP (tool integration). The article breaks down the layers of an agent's context, four typical failure scenarios (poisoning, distraction, confusion, conflict), and a context management approach through write, select, compress, isolate.

Anthropic

OpenAI

Google/DeepMind

Alibaba/Qwen

LangChain

Habr — хаб NLP27.07 · 09:05

Show more ↓