DGX Spark 主打在桌面端本机跑大模型：单机 128GB 统一记忆体，两台互连可达 256GB；以 ConnectX-7 连线提供 200Gbps 频宽。支援 NVFP4 后，可在维持精度下大幅降低记忆体占用并提升吞吐；例如在双机配置上，Qwen-235B 以 NVFP4 加上 speculative decoding，相较 FP8 可达最高 2.6 倍效能提升；量化到 NVFP4 亦可让记忆体用量约下降 40%，带来更好的多工与回应性。
开源合作也带来可量化收益：Llama.cpp 在 DGX Spark 跑 MoE 模型平均提升约 35%。在创作端，128GB 记忆体可于全精度执行如 GPT-OSS-120B 或 FLUX 2（约 90GB）等模型；新一代影音模型（如 LTX-2，NVFP8 权重）提升桌面端影片生成的可行性。DGX Spark 纳入 NVIDIA-Certified Systems 计划并进行测试，同时推出多个上手 playbook（如 Nemotron 3 Nano 30B、双机 FSDP+LoRA 微调至 70B），并宣布 Brev 的本机算力支援将于 CES 预览、预计 2026 年春季正式推出。
At CES 2026, NVIDIA’s latest DGX Spark software release pairs with updated models and open-source libraries to boost local AI performance across inference, training, and creator workflows, benefiting both DGX Spark and OEM GB10-based systems.
DGX Spark targets large-model work on the desktop with 128GB unified memory; two systems can link for 256GB. They connect via ConnectX-7 at 200Gbps for low-latency multi-node workloads. With NVFP4, next-gen models can cut memory footprint while increasing throughput: on a dual-system setup, Qwen-235B using NVFP4 plus speculative decoding delivers up to 2.6x higher performance than FP8, and NVFP4 quantization reduces memory use by about 40% while preserving accuracy and responsiveness.
Open-source collaboration adds measurable gains, including an average 35% uplift from Llama.cpp when running MoE models. For creators, DGX Spark can run large models like GPT-OSS-120B and FLUX 2 (about 90GB) at full precision, while newer video models (e.g., LTX-2 with NVFP8-optimized weights) improve desktop video generation. DGX Spark is joining the NVIDIA-Certified Systems program, new playbooks include Nemotron 3 Nano (30B) and two-node fine-tuning up to 70B (FSDP+LoRA), and Brev’s local-compute support is previewed at CES with official support planned for spring 2026.
