A LocalLLaMA user tested a 2.9bpw quantized DeepSeek V4.1 Flash 1017 model on APEX 2025, reporting about 1,000 tokens per second for prefill and 75 for decoding on a six-RTX 3090 system.
The LocalLLaMA poster said the model's performance on Terminal Bench 2.1 and AIME 2026 was similar to an unquantized version, and chose APEX 2025 to assess the 2.9bpw quant.
The test system had six RTX 3090 GPUs, 128GB of DDR4 memory and a Threadripper 3960X processor. The user said they ran the model with Project Maya.