DeepSeek V4.1 Flash: Long Context With ZDR Routing

DeepSeek V4.1 Flash is a fast, low-cost multimodal model on magicdoor.ai with a 1,048,576-token context window. It costs $0.22 per 1 million input tokens, $0.007 per 1 million cached input tokens, and $0.66 per 1 million output tokens.

Every request is pinned to Fireworks through Vercel AI Gateway with zero data retention enforced. If that exact route is unavailable or no longer qualifies for ZDR, the request fails closed. It never falls back to Together, DeepSeek direct, or another provider.

Open the DeepSeek V4.1 Flash model page for the current model summary.

Capabilities at Launch

The live Gateway metadata advertises vision, tool use, and reasoning support. magicdoor.ai enables image input and its normal tools while keeping unsupported file and code-interpreter surfaces disabled. The explicit reasoning control stays off until Gateway reasoning can be wired without inventing reasoning-effort semantics:

Switch to another supported model when a task needs PDFs, arbitrary files, code interpreter, or an explicit reasoning control.

Pricing

Billing uses the usage reported for the response. Fireworks reports cached input at $0.007 per million tokens when provider usage identifies cache hits; magicdoor.ai does not force a cache hit.

Examples:

• 100,000 input tokens plus 20,000 output tokens costs about $0.0352.
• 500,000 input tokens plus 100,000 output tokens costs about $0.176.

These are model-usage costs. magicdoor.ai has a $6 monthly base subscription that includes $1 in credits, with additional usage billed from your pay-as-you-go balance.

What Zero Data Retention Means Here

For this model, zero data retention applies to the external model-provider route. Both the ZDR requirement and the Fireworks allowlist are enforced on every request. A request returns an availability error rather than falling back to another provider.

This does not mean end-to-end encryption, and it does not disable magicdoor.ai's normal user-requested chat persistence. Saved chats continue to be stored so users can keep history, unless they use an ephemeral product flow.

The model advertises implicit prompt caching. Cached-input billing applies only when Fireworks reports a cache hit; magicdoor.ai does not force caching or relax the ZDR rule to obtain it.

Existing Chats and Links

DeepSeek V4 Flash 0731 is no longer an active picker entry. Existing chats, model-page links, stored IDs, and common old label variants resolve to DeepSeek V4.1 Flash. Its old GreenPT alternate was tied to the retired 0731 model and is disabled rather than presented as V4.1 support. DeepSeek V4.1 Flash now has its own optional GreenPT renewable route.

When to Use It

Choose DeepSeek V4.1 Flash for low-cost text and image tasks, large contexts, drafting, rewriting, summarization, tool-driven work, and a second model's perspective. Switch when the task requires a capability that is not enabled on this route.

Because magicdoor.ai preserves the conversation while you switch models, this remains one workflow rather than a collection of separate tools.

FAQ

How much does DeepSeek V4.1 Flash cost on magicdoor.ai?

$0.22 per 1 million input tokens, $0.007 per 1 million cached input tokens, and $0.66 per 1 million output tokens.

Does DeepSeek V4.1 Flash support search, files, or tools?

Image input, search, image generation/editing, Canvas, memory, Oggie, and tool calling are enabled. PDFs, arbitrary files, code interpreter, and the explicit reasoning control are not enabled.

Does DeepSeek V4.1 Flash use zero data retention?

Yes. Every request is pinned to Fireworks through Vercel AI Gateway with ZDR required and fails closed if that route is unavailable. This provider policy is separate from normal magicdoor.ai chat-history persistence.

Can old DeepSeek V4 Flash 0731 chats continue?

Yes. Old IDs and label variants resolve to DeepSeek V4.1 Flash, and legacy model-page URLs redirect to the new canonical page.