DeepSeek V4.1 Flash Operational

DeepSeek V4.1 Flash is an intermediate model under internal testing by DeepSeek. It uses a new model architecture, natively supports multimodality, and focuses on stronger capabilities, faster speed, and lower cost. Keep the base_url unchanged when calling the model, set the model name to deepseek-v4.1-flash-expires-on-0910, and note that each account is limited to 20 concurrent requests. Community testing has measured output speeds above 300 tokens/s.

New model architecture: An intermediate version using a completely new architecture, paving the way for the subsequent official release.
Native multimodality: Natively supports image understanding without an external vision module.
High-speed inference: Community testing has measured output at 300~400 tokens/s, with a peak of 507 tokens/s.
Enhanced capabilities: Stronger capabilities than V4 Flash in the same category, with better performance on coding and generation tasks.

New model architecture: Following DeepSeek's established approach, with the V4 series based on a mixture-of-experts (MoE) architecture, industry observers speculate that V4.1 Flash redesigns routing and layer structures on top of MoE, using fewer active parameters to achieve higher throughput.
Native multimodality: Unlike V4 Flash, which offers a separate Vision version, V4.1 Flash integrates visual encoding directly into base-model training, aligning text and images in the same representation space and avoiding the latency and information loss of cross-module concatenation.
High-speed inference: High throughput typically comes from sparse MoE activation, where only some experts are computed for each token, operator-level inference optimization, and cache reuse. Whether these techniques are used specifically remains to be disclosed in the official release.

Register an account: Visit the DeepSeek Open Platform at https://platform.deepseek.com/, then register and complete real-name verification.
Get a key: Create and copy your API Key on the "API keys" page.
Check your balance: Make sure the account has been funded. Pricing is the same as deepseek-v4-flash.
Configure the endpoint: Keep base_url unchanged at https://api.deepseek.com.
Set the model name: Set the model parameter in the request to deepseek-v4.1-flash-expires-on-0910.
Make a call: Send requests using the standard OpenAI-compatible format to try the new model.
Mind the concurrency limit: Each account is limited to 20 concurrent requests, so control request volume for high-frequency workloads.

Exceptional speed: Community testing generally measures output at 300~400 tokens/s, peaking at 507 tokens/s, making interactions feel nearly instantaneous.
Native multimodality: Visual capabilities are integrated directly into the base architecture, eliminating the need for a separate vision version as with V4 Flash.
Same price, effectively lower cost: Capabilities are upgraded while pricing remains at the V4 Flash level, with cached input during off-peak hours as low as 0.05 yuan per million tokens.
Renewed architecture: Uses a completely new model structure and serves as a preview of DeepSeek's next technical direction, leaving more room for future upgrades.
Zero-cost integration: Keep base_url unchanged and change only the model name, enabling a seamless switch for existing users.

Real-time conversational assistants: Streaming output at 300+ t/s makes chat and customer service scenarios feel virtually wait-free.
AI coding assistance: Code completion and snippet generation are highly latency-sensitive, and fast output can be integrated seamlessly into IDEs.
Vision understanding applications: Native multimodality suits image-text tasks such as screenshot Q&A, document image analysis, and UI review.
Lightweight Agent execution layer: Agents require frequent sequential model calls, making low cost and high throughput suitable for execution and routing tasks.
Batch content production: Pipeline tasks such as copy rewriting, summarization, and translation can minimize costs through low unit prices and cache hits.