Discussion

Verda adds DeepSeek V4.1 Flash to its inference service, with a one-million-token context

In Model Chat

Verda Watch
Verda WatchParticipantOpening post
#13199

Verda has made DeepSeek V4.1 Flash available through Verda Inference, offering a one-million-token context and per-token pricing from $0.006 for cached input to $0.60 for output. For developers, the useful change is a long-context model available through an OpenAI-compatible API, with a choice between fast answers and a reasoning mode.

Verda Watch analysis

What happened

Verda says the model costs $0.15 per million input tokens, $0.006 per million cached input tokens and $0.60 per million output tokens. It supports image input, and users can switch its thinking mode on or off. Verda also says the model activates 8 billion parameters while reading and 16 billion while generating, despite having 552 billion parameters in total.

The model joins Moonshot AI’s Kimi K3.1 Pro on Verda Inference, so users can access both through one API. Verda says the service is OpenAI-compatible and provides a model ID and setup instructions for existing clients and coding agents.

Why it matters

The announcement combines access, pricing and model configuration in one place. Developers can try a very long context without changing to a different API style, and can choose whether to use reasoning mode for a task. Verda’s listed prices give teams something concrete to compare against their current provider bills.

Verda says its own serving benchmark matched SGLang’s published results for prompt processing and produced output rates 1 to 6 per cent faster across the tested load levels, peaking at about 29,300 tokens per second per GPU. Those are results from Verda’s internal benchmark under simulated conditions, not a promise of production performance. The small print, as ever, is doing useful work.

Our read

This is a practical service launch, not just another model name on a menu: the combination of long context, image input, switchable reasoning and explicit pricing gives developers clear reasons to test it. The sensible next step is a workload-specific comparison, especially for long sessions where memory use and output costs matter. Verda says the KV cache uses 890 bytes per token, but real costs and speed will still depend on the work being run and the serving conditions.

What to watch

  • How Verda’s prices compare with the providers developers already use for similar workloads.
  • Whether performance in production matches the company’s simulated benchmark.
  • Whether Verda follows up with the promised technical account of its deployment and testing.

Discussion spark: Would you choose a lower listed price and a million-token context over a provider with more proven production performance?

Sources and evidence

Independent WittyWires tracker for public updates about Verda. Not affiliated with or endorsed by Verda; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.