DeepSeek V4.1 Flash launches with native multimodal vision and a sharp price cut

On September 10, 2026, DeepSeek released V4.1 Flash, a model with native multimodal visual understanding, trained with a new pre-training approach and a larger reinforcement-learning post-training stage. Pricing is markedly cheaper than the previous generation: in off-peak hours input token costs drop to ¥0.02 per million when cache-hitting and ¥1 per million when cache-missing, with output at ¥4 per million, and peak hours cost roughly double the off-peak rate. Community tests suggest throughput of roughly 300-355 tokens per second, occasionally topping 500. This article explains the feature set, the pricing logic, and who it suits.

Background

• DeepSeek has built a reputation for releasing capable models at disruptive prices, and V4.1 Flash continues that strategy with an emphasis on multimodal input.
• Native multimodal vision means the model can accept images alongside text without a separate vision adapter, which simplifies pipelines for document, screenshot, and diagram work.
• The tiered pricing incentivizes developers to push heavy traffic to off-peak hours, a model-level approach to managing capacity.

Key facts

Item	Detail
Model	DeepSeek V4.1 Flash
Multimodal	Native visual understanding
Off-peak output price	¥4 / million tokens
Off-peak cache-hit input	¥0.02 / million tokens
Off-peak cache-miss input	¥1 / million tokens
Peak pricing	~2x off-peak
Community throughput	~300-355 tok/s (peaks >500)

Highlights

Native multimodal without an adapter

The main feature advance is accepting images natively, which removes the friction of routing visual work through a separate encoder or keeping two models in sync. For teams processing PDFs, screenshots, product photos, and diagrams, a single model simplifies deployment and cost accounting. The image below shows a schematic of multimodal input to a model:

Caption: Native multimodal models take images and text into one pipeline, which is the core simplification V4.1 Flash brings.

The off-peak pricing lever

The sharply discounted off-peak rates are a pricing design that moves cost-sensitive batch work to quieter hours. Teams that can schedule bulk processing, evaluations, or nightly indexing around the cheaper window can cut their bill substantially, though interactive users pay the premium rate during the day.

Throughput and latency signals

Community informal tests reporting 300-355 tokens per second, with peaks above 500, suggest a fast inference experience for agents and high-volume consumers, although these figures come from community benches rather than official documentation.

Industry positioning & impact

DeepSeek V4.1 Flash continues the pattern that has defined DeepSeek's market role: shipping competitive capabilities at a fraction of rival pricing to pressure the entire AI pricing landscape. The native multimodal addition pushes that story beyond text, while the off-peak tiering shows DeepSeek managing capacity through price incentives instead of hard limits. Each aggressive launch forces larger vendors to justify premiums and keeps the cost of AI falling globally, which benefits developers and end users.

The competitive and strategic read is significant for the open versus proprietary model debate. A fast, cheap, multimodal model erodes the pricing moat that proprietary vendors rely on, and its release follows the broader theme of cost pressure visible across recent pricing and infrastructure moves. The real tests are reliability under production load, quality of the multimodal understanding on real documents, and whether the throughput figures hold at scale. DeepSeek's official channels remain the authoritative source for specs and pricing, and the community will watch for independent quality comparisons before drawing final conclusions about model quality versus price.

Related reading

For how cheap, capable models reshape the competitive field, see the best AI coding tools comparison 2026 and the AI model outages and concentration risk pieces. On the capital race behind such pricing, the DeepSeek Huawei Ascend build-out story adds useful context on supply.

References

The launch and pricing details were reported by Chinese tech-media in coverage such as Toutiao/Sci-Tech Innovation Board Daily and 21 Business News, with throughput discussed by developers and summarized in independent reviews like AI Tools Review.

Buying advice & audience

If you are searching "DeepSeek V4.1 Flash", "DeepSeek pricing 2026", or "multimodal AI cheap model", the deciding factor is your traffic pattern. Teams with heavy, schedulable batch work should exploit the off-peak cache-hit and ¥4-输出 pricing to cut costs sharply, while interactive users should weigh the day-rate premium against the model's native multimodal quality. Integrating teams that process documents, screenshots, and diagrams will gain the most from the native vision. Rely on the official pricing page as the source of truth for tiered rates and cache behavior. Rising interest in "DeepSeek multimodal", "cheap vision LLM", and "DeepSeek V4.1 Flash review" suggests sustained demand, and this article covers features, pricing, and fit so you can decide without overpaying.

FAQ

Is DeepSeek V4.1 Flash worth using?

For cost-sensitive multimodal work that can be batched into off-peak hours, yes; interactive users should test the native vision quality and factor in the day-rate premium before committing.

How much does DeepSeek V4.1 Flash cost?

Off-peak output is ¥4 per million tokens, cache-hit input ¥0.02 and cache-miss ¥1 per million, with peak rates roughly double; official pricing is authoritative.

What does native multimodal mean here?

The model accepts images directly in its input without a separate vision adapter, simplifying pipelines for documents, screenshots, and diagrams and easing deployment.

How fast is it in practice?

Community informal tests report roughly 300-355 tokens per second with peaks above 500, though these are community benches rather than official figures.

Who should use DeepSeek V4.1 Flash?

Teams doing batch-heavy, cost-sensitive processing with multimedia input benefit most; interactive or latency-critical users should validate day-rate cost and throughput on their workloads.