DeepSeek V4 Flash: Local Performance and Cost Efficiency

Performance and Benchmark Context

DeepSeek V4 Flash demonstrates significant efficiency gains through post-training, outperforming larger models like GLM 5.2 and approaching the performance of Opus 4.8. The model shows a marked improvement in agentic coding tasks, with its score on the DeepSeek agent harness jumping from 7% to 54%. The author emphasizes that benchmark results are highly dependent on the harness used; for example, optimizing a harness for ARC-AG can improve scores from 13% to 40%. Consequently, DeepSeek's internal benchmarks remain the most reliable indicator of model improvement, as they are shielded from public data contamination.

Local Execution and Hardware Requirements

Running DeepSeek V4 Flash locally requires substantial VRAM due to the model size and context window requirements. For 4-bit precision, approximately 168GB of VRAM is necessary, while 3-bit quantization requires roughly 110GB. To optimize memory usage, developers can utilize the Dwarf Star engine, which enables SSD offloading and mixed-precision techniques (e.g., keeping some weights in 4-bit while others are in 2-bit). This allows the model to run on systems with 128GB of VRAM while leaving sufficient headroom for the KV cache. In a test setup using two NVIDIA DGX Sparks, the model achieved speeds of 20 to 30 tokens per second.

Agentic Coding Capabilities

DeepSeek V4 Flash is positioned as a workhorse model for agentic tasks rather than a general-purpose frontier model. It supports a 1M token context window and integrates with open-source harnesses like Open Code. In demonstrations, the model successfully generated a functional legendary Pokémon encyclopedia and an International Space Center tracker, maintaining a coherent chain-of-thought process. The model is currently text-only and lacks native multimodal or vision capabilities.