A user has benchmarked the DeepSeek V4.1 Flash 1017 EXL3 2.9bpw model, finding its performance comparable to state-of-the-art models from six months prior. Using the APEX 2025 dataset as a benchmark, the user achieved approximately 1000 tokens per second for prefill and 75 tokens per second for decoding on a system with six 3090 GPUs. The user shared these results to assist others in benchmarking their own setups. AI
IMPACT Provides performance insights for a specific model quantization, aiding users in evaluating its capabilities.
RANK_REASON User-generated benchmark results for a specific model variant. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →