OpenCode has announced that its DeepSeek V4.1 Flash model is now available for free. This move follows a series of adjustments to its usage policies, including temporary and permanent credit increases for the model within the OpenCode Go plan. The free tier of DeepSeek V4.1 Flash will support a 200K context window, a reduction from the 1 million token context version previously available.
DeepSeek V4.1 Flash: A Recap
Officially launched on September 10, 2026, DeepSeek V4.1 Flash is a native multimodal MoE (Mixture of Experts) model. It boasts a backbone parameter scale of 552B, complemented by 196B Engram conditional memory parameters. The model features a causal encoder-decoder architecture, with the first 20 of its 40 language backbone layers acting as an encoder and the latter 20 as a decoder. This design significantly reduces the Prefill computation, bringing it close to half the original amount.
On the official DeepSeek platform, there is no free API tier for this model. Instead, usage is typically billed on a pay-as-you-go basis with pre-paid top-ups. OpenCode’s decision to offer a free tier, even with a reduced context window, marks a significant step in making advanced AI models more accessible.
Context Window and Performance
The free version of DeepSeek V4.1 Flash offers a 200K context window. While this is a substantial context length, it’s a decrease from the 1M context version that was previously part of the OpenCode Go package. The reduction in context window size for the free tier likely aims to balance accessibility with resource management.
User feedback regarding the model’s performance, particularly its tokens per second (TPS) or generation speed, has been a significant point of discussion in the community. Reports indicate considerable fluctuations in speed. Some users on platforms like ollama cloud have observed peak speeds reaching around 490 tokens/s. However, consistency remains an issue, with speeds reportedly dropping to as low as 50-60 tokens/s during peak hours, such as midday to afternoon.
Within the OpenCode Go plan itself, users have also reported TPS falling below 100. Further community testing has shown that the speed performance degrades as the context length increases. Some users noted that with contexts exceeding 100K, the TPS dropped below 100, and at the 200K mark, it hovered around 50 tokens/s. This suggests that while the model can handle large contexts, its efficiency decreases significantly with longer inputs.
OpenCode’s Evolving Strategy
This free offering is a continuation of OpenCode’s “Operation Cheepseek” initiative. Earlier in September, OpenCode had temporarily increased the usage credit for DeepSeek V4.1 Flash within its OpenCode Go package from $15 to $60. This was subsequently adjusted to $80 on October 12th. With the introduction of the free tier, calls to the model on the OpenCode platform no longer consume these credits, making it a cost-effective option for many users.
This strategic move by OpenCode aims to broaden the adoption and testing of DeepSeek V4.1 Flash, allowing a wider audience to experience its capabilities without the initial financial barrier. The performance fluctuations are a common challenge with large, complex models, and continued optimization by OpenCode and the DeepSeek team is expected.
Related Developments
This announcement is complemented by ongoing developments in the DeepSeek ecosystem. Recent updates include:
- OpenCode’s previous announcement regarding the permanent $60 credit for DeepSeek V4.1 Flash.
- The official release of DeepSeek V4.1 Flash, highlighting its advancements over V4 Pro, including native multimodal visual understanding and significant price reductions of up to 60%.
- Updates to DeepSeek Harness (version 0.1.5), which now supports the DeepSeek V4.1 Flash model and introduces features like file uploading and sidebar previews.
These related developments underscore a dynamic environment of innovation and user-centric improvements surrounding the DeepSeek models.
