#DeepSeek V4
5 articles with this tag
Crusoe Keeps 4-Bit Weights to Cut LoRA Memory by 75%
Crusoe shows how storing frozen base weights in 4-bit and dequantizing in-kernel cuts LoRA memory by up to 75% without accuracy loss.
Qwen3 8-27B Model Challenges Frontier AI Performance
The Qwen3 8-27B model has achieved impressive results on the Artificial Analysis Benchmark, demonstrating performance comparable to leading AI models like DeepSeek V4 and GPT-5.6 Luna Max. This development suggests that near-frontier AI capabilities are becoming more accessible.
Liang Wenfeng's DeepSeek: $5.6M Training Budget, $45B Valuation
DeepSeek-V3 cost $5.6 million to train. V4-Pro charges $1.74 per million tokens. And Liang Wenfeng is now raising $10 billion at a $45 billion valuation — here is what we know about DeepSeek's finances.
DeepSeek V4 vs. Opus: Ahmad Awais on AI Coding Taste
Ahmad Awais discusses how AI coding agents can learn 'coding taste' to outperform generic models, focusing on the difference between functional code and good design.
DeepSeek-V4: Million-Token Context is a Serving Problem
DeepSeek-V4's million-token context window presents an inference systems challenge, demanding sophisticated cache management and serving strategies to unlock its potential.
