DeepSeek V4 Flash hits public beta at $0.14/million tokens, beats V4 Pro on agents

DeepSeek moved its V4 Flash model into public beta on July 31, and the numbers are worth your attention — especially if you’re building agents on a budget. The 284B-parameter MoE (with 13B active per token) now beats the pricier V4 Pro on every agentic benchmark the company published, while costing a third as much. Input pricing lands at $0.14 per million tokens on a cache miss and $0.0028 on a hit, with output at $0.28 — roughly a third of V4 Pro’s $0.87. That’s the kind of pricing that lets seed-stage startups run agent loops without a GPU budget.

The model card is explicit: this isn’t a new architecture. What changed is post-training. The April preview architecture remains intact — 1 shared expert and 256 routed experts per MoE layer, hybrid attention mixing Compressed Sparse Attention with Heavily Compressed Attention, and those Manifold-Constrained Hyper-Connections replacing standard residuals. What’s new is the DSpark speculative decoding module, which ships attached to the checkpoint and delivers 60–85% faster per-user generation versus the baseline. DeepSeek’s vLLM example serves this on a single 4×GB300 node.

Self-hosters face a steeper climb. The weights are MIT-licensed and ungated, which is genuinely refreshing, but every expert stays resident in memory even though only 13B activate per token. Unsloth’s dynamic GGUFs put the lossless 8-bit build at 162 GB and a 3-bit build at 103 GB, needing roughly 110 GB of combined RAM and VRAM. That’s one well-specced workstation at aggressive quantization, or a proper serving cluster for mid-size enterprises. Not trivial, but the licensing makes it possible.

I’d pump the brakes on the benchmark euphoria, though. DeepSeek ran its agent evaluations using the minimal mode of DeepSeek Harness — which hasn’t been released publicly — and two of the test sets, DSBench-FullStack and DSBench-Hard, are internal. Agent scores are notoriously harness-sensitive, so independent runs may tell a different story. The 1M-token context window and native Responses API support (plus Codex adaptation) make this a compelling package, but until third-party evals land, those chart-topping numbers are vendor claims on vendor turf. Run your own tests before betting production workloads on them.