OpenCode has made an exciting announcement—DeepSeek V4.1Flash is now accessible for free on its platform. Users can now invoke this model without it counting against their subscription quota. Based on real-world tests, this complimentary version supports a 200K context window, which, while impressive, is smaller than the 1M context window version previously available on the platform. This change represents a significant progression in OpenCode's ongoing "Operation Cheepseek" initiative. Before this development, OpenCode had already enhanced the usage quota for DeepSeek V4.1Flash twice within its subscription offerings.

Currently, the official DeepSeek platform does not provide a free tier for its API. Instead, it operates on a prepaid, pay-as-you-go basis. Launched on September 10, 2026, DeepSeek V4.1Flash is a cutting-edge native multimodal Mixture of Experts (MoE) model. It boasts a primary parameter scale of 552 billion, complemented by an additional 196 billion Engram conditional memory parameters. This unique architecture is designed to reduce the Prefill computational load by nearly half. However, users may notice some variability in the model's generation speed, with significant differences in output speed depending on the platform and time of use. Specifically, when the context size surpasses 100K, the generation speed dips below 100 tokens per second; at 200K, it averages around 50 tokens per second.