product
DeepSeek v4 Price Change Redefines the Enterprise API Math Against OpenAI and Anthropic
The new DeepSeek v4 pricing adjustment challenges the ultra-cheap cost-performance ratio that made the Chinese AI startup famous. Here is what it means.
Company
DeepSeek
Chinese artificial intelligence company. Ultrathink has published 13 stories on DeepSeek. Latest, on : DeepSeek v4 Price Change Redefines the Enterprise API Math Against OpenAI and Anthropic.
ultrathink.ai
product
How the new DeepSeek Harness democratizes enterprise-grade LLM evaluation and training replication
DeepSeek Harness open-sources the evaluation infrastructure behind the world's most efficient AI models, empowering developers to replicate LLM testing.
ultrathink.ai
product
Alibaba leaks Qwen 3 MoE model on Hugging Face with 2.4 trillion parameters
Alibaba leaks Qwen3.8-2.4T-A95B-FP8 on Hugging Face, signaling a massive 2.4 trillion parameter MoE model to rival Meta's Llama open-source dominance.
ultrathink.ai
breaking
DeepSeek V4 Pro Appears on OpenRouter: What to Expect from the Next-Gen Model
DeepSeek V4 Pro has silently appeared on OpenRouter. Read our analysis of DeepSeek's next-gen model and what it means for API pricing.
ultrathink.ai
analysis
Why Cursor's Mixture-of-Kittens MoE Megakernel Solves Blackwell's Hidden CPU Bottleneck
The new Mixture-of-Kittens MoE megakernel by Cursor bypasses Blackwell Grace CPU bottlenecks, boosting MXFP8 throughput by 2.37x on NVL72 hardware.
ultrathink.ai
product
How Llama.cpp DeepSeek-V4 Support Delivers Massive Local Inference Performance Gains
Get native llama.cpp DeepSeek-V4 support with release b10228. Run DeepSeek-V4 with Multi-Token Prediction (MTP) locally on consumer hardware today.
ultrathink.ai
breaking
DeepSeek V4 Flash: Benchmark Data Reveals Next-Gen Price-to-Performance King
DeepSeek V4 Flash benchmark data on Artificial Analysis reveals a new leader in low-cost, high-speed AI inference, pressuring OpenAI and Anthropic.
ultrathink.ai
analysis
How US Startups Became Dependent on Chinese AI Models—and Why a Ban Hurts Silicon Valley
US AI startups are lobbying against a ban on Chinese open-weight AI. Restricting models like Qwen or DeepSeek could cripple American tech innovation.
ultrathink.ai
analysis
How Moonshot AI Built a Frontier Competitor Without Just Copying Anthropic’s Fable
Analyze Kimi K3 performance and why experts say Moonshot AI did not rely on distilling Anthropic's Fable. Explore the true scaling breakthroughs.
ultrathink.ai
analysis
Why the Treasury’s Threat to Sanction Chinese Open-Source AI Will Fail
US threats of Chinese open-source AI sanctions target IP theft, but enforcing blocks on platforms like Hugging Face presents massive technical hurdles.
ultrathink.ai
product
How the LangChain reasoning effort parameter simplifies managing compute costs for reasoning models
The new LangChain reasoning effort parameter in core 1.5.0 standardizes how developers control compute and cost for reasoning LLMs.
ultrathink.ai
product
vLLM's Wide Expert Parallelism Makes DeepSeek Inference 10x More Efficient at Scale
vLLM achieves 2,200 tokens per second per H200 GPU serving DeepSeek using wide expert parallelism. A technical breakthrough for MoE model inference costs.
ultrathink.ai
product
NVIDIA's 72GB Desktop GPU Makes Running Large Language Models Locally Practical
NVIDIA RTX PRO 5000 72GB Blackwell GPU is now available, giving AI developers enough VRAM to run 70B+ parameter models locally. Here's what it means for agentic AI.