DeepSeek V4 matches giants, still pricey

DeepSeek V4 is a real step forward for open-weight long-context models, but it also reinforces the uncomfortable truth that “open source” does not mean “easy to run.” The architecture looks designed for agent workloads first, local hobbyist inference second.

• –The release centers on V4-Pro and V4-Flash, with 1M-token context and MoE designs aimed at cutting compute and KV-cache cost
• –Official and third-party writeups place it near Claude Opus, Gemini, and GPT-5-class models on several coding and agent benchmarks
• –The practical bottleneck is hardware: even with efficiency gains, the model sizes still push most users toward quantization, server-grade GPUs, or hosted inference
• –That makes the release most interesting as an engineering signal: long-context agent models are getting cheaper to serve, but not yet cheap enough to feel local
• –For developers, the bigger story is not “can I run it on my laptop?” but “can my stack handle million-token workflows without collapsing?”