CONFIRMED: the two-day experiment became an official product

DeepSeek confirmed the launch of V4.1 Flash on September 10, 2026, closing the limited test phase that had placed a temporary identifier inside its API. The model now appears in the company’s official documentation as the smallest member of its new architecture family. The central change is not just a name: Flash now has native visual understanding instead of relying on the experimental Vision Exp model, and it was designed to raise the capability ceiling, speed up inference, and increase throughput as the family scales.

The architecture is the most important part of the story

The official presentation describes a 552 billion parameter MoE with a Causal Encoder Decoder architecture. According to DeepSeek, only 8 billion parameters are active on input and 16 billion on output. The company also says the KV cache uses one quarter of the HBM memory and one eighth of the SSD storage required by the previous generation. That is a structural efficiency claim, not just a commercial price cut, and it could reduce the cost of agents that repeat context across many calls.

What actually shipped

The API documentation lists a 1 million token context window, a maximum output of 384,000 tokens, thinking and non-thinking modes, tool calls, the Responses API, Anthropic API compatibility, and vision support. The recommended name for the current version is deepseek-flash. The previous identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to the new model, an important decision for teams that already have a production integration and do not want to change everything today.

Open weights broaden the bet

On the same day, the official Hugging Face repository began distributing the weights under the MIT license, with dozens of safetensors shards and inference instructions. That does not turn the model into a simple installation for every company: its total size still requires substantial infrastructure and software support. But it creates a second route beyond the API, relevant to organizations that need data control, customization, or cost negotiation at scale.

Ambitious benchmarks, but they are still DeepSeek’s own numbers

The official table puts V4.1 Flash at 90.9 on GPQA Diamond, 3,471 on Codeforces, 90.6 on Terminal-Bench 2.1, and 74.2 on DeepSWE v1.1. On tool-using tasks, the company reports 63.9 on HLE, 31.8 on Agents’ Last Exam, and 89.6 on BabyVision. That points to a much more aggressive bet on coding agents and multimodal tasks than the previous Flash. The caveat is substantial: this release does not provide an independent evaluation confirming the advantage under real workloads, total cost, or long-term stability. A lab benchmark is evidence, not a production verdict.

The bill gets smaller and V4 Pro loses ground

The official pricing page lists Flash at $0.003 per million cached input tokens, $0.15 for uncached input, and $0.60 for output outside peak hours. Peak rates are double. DeepSeek also says that after noon Beijing time on September 14, all requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the Flash rate until V4.1 Pro launches. In practice, the provider is turning the cheaper model into the default path for a large share of demand.

What changes for Brazilian companies

For a small business, the most interesting gain is architectural. One endpoint can analyze documents and images, call tools, and return long answers, reducing the need to stitch together a text model and a separate vision service. That fits catalogs, support with attachments, creative checks, contract reading, and coding agents. The dollar price helps, but it does not erase exchange-rate exposure, observability, concurrency limits, or fallback requirements. The safe migration is to test real prompts before accepting the automatic Pro routing, especially on tasks that require vision and consistency across many turns.

MaxAssistant’s read

The story moved from strong signal to confirmed, and it now has two fronts: a cheaper API and open weights. DeepSeek did more than launch a fast Flash: it put multimodality, tools, memory compression, and a Pro replacement strategy in the same package. My read is that the company is compressing the gap between its own API tiers to win volume, not merely chasing a benchmark table. If the numbers hold outside the official harness, V4.1 Flash could be highly competitive for high-volume agents. If they do not, the low price will be the only differentiator, and price alone cannot save an automation that makes mistakes.

Sources

DeepSeek, V4.1 Flash announcement: https://deepseek.com/news/deepseek-v4-1-flash | DeepSeek API Docs, Change Log: https://api-docs.deepseek.com/updates/ | DeepSeek API Docs, Models & Pricing: https://api-docs.deepseek.com/quick_start/pricing | Hugging Face, official model repository: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash | ccleaks, independent coverage of the launch and open weights: https://ccleaks.com/news/deepseek-v4-1-flash-open-weights-sep-2026 | TechNode, limited beta coverage: https://technode.com/2026/09/09/deepseek-v4-1-flash-multimodal-limited-beta