Skip to main content

DeepSeek's New Flash Model Lands on Qwen, and It's Cheap

DeepSeek-V4.1-Flash Arrives on Qwen AI Platform

DeepSeek has quietly dropped its latest lightweight flagship, DeepSeek-V4.1-Flash, onto the Qwen AI platform. The move opens up both API access and a Token Plan, giving developers two ways to tap into the model. You can wire it into your own systems via standard APIs, or use the Token Plan directly inside tools like Qoder, Qwen APP, and Codex for coding, documentation, visual understanding, and agent tasks.

Image

What's Under the Hood?

At its core, V4.1-Flash uses a Mixture-of-Experts (MoE) architecture with a hefty 552 billion total parameters. But here's the twist: it only activates about 8 billion parameters on the input side and roughly 16 billion on the output. That asymmetric Causal-Encoder-Decoder design keeps things fast without sacrificing muscle.

The model natively understands both text and images, and it can handle a maximum context length of 1 million tokens—with output stretching up to about 393K tokens. According to DeepSeek, it's made significant leaps over its predecessor on several agent and code benchmarks, all while keeping throughput high and latency low thanks to those modest activation numbers.

The Cost Story: Cheaper, Faster, Leaner

Pricing is where things get interesting. A new generation of cache compression slashes KV Cache demands—down to a quarter of the previous generation for HBM and just one-eighth for SSD storage. That means long contexts and multi-turn tool calls won't gobble up resources like before.

On the Qwen page, you'll find time-based pricing:

  • Off-peak hours: 1 yuan per million input tokens, 4 yuan per million output tokens.
  • Peak hours: 2 yuan and 8 yuan respectively.

Rate limits sit at 15K RPM and 1M TPM, covering prefix completion, function calls, caching, structured output, batch tasks, and online search.

Deployment and Ecosystem

Alibaba Cloud's BaiLian has teamed up with the vLLM open-source community to get everything adapted. The model is now accessible on both the China site and the international site, spanning regions like Beijing, Singapore, and global spots including the US, Germany, Japan, and Hong Kong.

According to BaiLian's documentation, V4.1-Flash supports multi-turn conversations, function calls, online search, context caching, and structured output—with a max_tokens limit around 393,216. That makes it a solid fit for enterprises looking to migrate tasks like long document parsing, customer service knowledge bases, code repository Q&A, and multimodal order review onto a single interface.

What the Industry Is Saying

Experts see this as a smart move: by baking the "large parameters, small activation, strong caching" combo into the Qwen ecosystem, V4.1-Flash could lower the trial-and-error costs for long-context agents. For smaller teams, the low off-peak pricing plus the Token Plan subscription means more flexible batch inference and prototype validation—without breaking the bank.

Key Points

  • DeepSeek-V4.1-Flash is now live on Qwen AI platform with API and Token Plan.
  • MoE architecture with 552B total parameters, ~8B input and ~16B output activation.
  • 1 million token context, native text and image understanding.
  • Cache compression cuts HBM demand to 1/4 and SSD to 1/8 of previous gen.
  • Pricing: 1 yuan/million input tokens off-peak, 4 yuan/million output; double during peak.
  • Available via Alibaba Cloud BaiLian in multiple regions, with vLLM adaptation.