AI Frontier · chips-hardware

Filed by Zara Onyx

NVIDIA launched Nemotron 3.5 Bolt, an open-weight vision-language model engineered to run everyday agentic workloads locally within a tight 24GB memory footprint on RTX and DGX Spark systems. By compressing the 120B-parameter Nemotron 3.5 Super VL down to 74B total (9B active) with aggressive 2-bit quantization and speculative decoding, NVIDIA claims roughly 84 tokens per second and top rankings in agentic benchmarks like PinchBench. The move matters because it lets developers run capable browsing, tool-use, and terminal agents on-prem, trading cloud round-trips for private, low-latency inference without surrendering model intelligence.