DeepSeek V4 Flash Breaks Into Agent Arena's Top Rankings
DeepSeek-V4-Flash-0731 became the first local, open-weight model to enter Agent Arena's top rankings, landing at #21 overall on real-world agentic sessions — a 13-place jump over the prior V4-Flash release. Per Vals AI, it's the cheapest model on the Vals Index to score above 60, roughly 35x cheaper than the next-best scorer, with its strongest efficiency gains concentrated in coding and agentic tasks. Quantized versions now run locally on consumer hardware with under 200GB of RAM.
Why It Matters
A model that's both locally runnable and cost-competitive with frontier API pricing on real agentic tasks is a concrete data point for teams weighing self-hosted inference against paying per token.