DeepSeek Releases V4 Flash 0731: 1M-Token MoE Model at $0.14/M Input Tokens

DeepSeek has released DeepSeek V4 Flash 0731, a sparse mixture-of-experts (MoE) model built for coding, reasoning, and agent workflows. The model is now available through OpenRouter's API under the identifier deepseek/deepseek-v4-flash-0731.

Architecture and Specs

V4 Flash 0731 uses a sparse MoE design with 13B active parameters drawn from a 284B total parameter pool — a ratio consistent with DeepSeek's prior efficiency-focused releases, where only a fraction of total weights activate per token to reduce inference cost.

The model ships with a 1,049K token context window (roughly 1 million tokens), positioning it among the largest-context models currently on the market. DeepSeek describes this as a "re-post-trained revision," indicating the underlying base model has undergone an additional post-training pass rather than a full pretraining cycle from scratch. Modality is limited to text-to-text; there is no indication of image, audio, or video input support.

Pricing

OpenRouter lists pricing at $0.14 per million input tokens and $0.28 per million output tokens. That places V4 Flash 0731 in the low-cost tier relative to frontier models from OpenAI, Anthropic, and Google DeepMind, consistent with DeepSeek's established strategy of undercutting Western labs on price.

What's Not Yet Known

DeepSeek has not published benchmark scores for V4 Flash 0731 alongside this release, and no independent evaluations have surfaced yet. The company has also not disclosed a training data cutoff date or detailed architecture paper for this specific checkpoint. Claims about performance on coding, reasoning, or agent tasks are, at this stage, DeepSeek's own positioning rather than externally verified results — read them as claims until third-party benchmarks appear.

What This Means

V4 Flash 0731 continues DeepSeek's pattern of shipping sparse MoE models that activate a small fraction of total parameters to keep inference costs low while retaining a large total capacity for knowledge and reasoning. The 1M-token context window is the headline spec here — it's large enough to handle full codebases, long agent trajectories, or extended multi-document reasoning tasks in a single pass, which is likely why DeepSeek is marketing it specifically for coding and agent workflows.

The pricing — a fraction of a cent per thousand tokens — keeps DeepSeek competitive against both Chinese rivals (Alibaba Qwen, Moonshot AI, Zhipu AI) and Western frontier labs. Until independent benchmarks are published, developers should treat this as an early-access, cost-efficient option worth testing against existing workloads rather than a confirmed leader on any specific task. The lack of a disclosed training cutoff and benchmark data is a gap worth watching for in follow-up documentation from DeepSeek.