Not available in English yet
DeepSeek V4 在家用小主机跑到 22-28 字/秒,本地大模型这事可能真要开始能用了
Related Reading
More on #DeepSeek
GLM Beats DeepSeek on Two GPUs — Chinese Open-Source Stops Compromising
On two NVIDIA DGX Sparks, GLM-5.3 Flash beat DeepSeek V4 Flash on HumanEval (97% vs 94.5%). GLM ran 30% slower with one-quarter the context.
DeepSeek Hits 67 token/s on Two $9K Mini Boxes — Local LLM Floor Is Caving In
Reddit user hit 67-84 token/s on DeepSeek V4 Flash with a 1M-token context window on two ~$9K NVIDIA DGX Sparks. The local-LLM cost barrier is collaps
Community Qwen3.8 Quantization Saves 30GB — Local LLM Bar Drops Again
Community dev agentionai ships a custom quantized Qwen3.8-Flash-Next, 20–30GB smaller than mainstream versions at comparable quality. Local LLM bar dr
Speculative decoding is becoming standard — open-source LLMs now predict ahead
Reddit users spotted speculative decoding working on local GPUs—AI instantly outputting phrases via MTP. The local inference cost curve is being quiet
Maxing AI 'Thinking Depth' Hurts Results — DGX Spark Local Test Warns Enterprises
A Reddit developer tested DeepSeek/Qwen on four DGX Sparks: 'deep thinking' mode lowers scores and doubles runtime — a direct cost warning for AI infe
DeepSeek's Agent Framework Exposes Five Keys as Chinese Devs Crack the Source
DeepSeek Agent framework has 5 event-dispatch modes decoded in an 8,000-word Chinese source tutorial - a shift from API calls to source-reading.