DeepSeek ships V4.1-Flash on its API with native multimodal and new pricing

DeepSeek published DeepSeek-V4.1-Flash, a multimodal mixture-of-experts model on the API as deepseek-flash with up to 1 million tokens of context and a posted off-peak cached-input price of $0.003 per million tokens.

DeepSeek put a cheap, long-context image-and-text model on its own price page, including a $0.003 off-peak cached-input rate. Cross-lab “beats GPT…” lines stay the company’s claims.

DeepSeek’s news page and API docs published DeepSeek-V4.1-Flash. The model name is deepseek-flash. It is natively multimodal, meaning it can take images as well as text. Older names still route to V4.1-Flash at Flash pricing — see Sources. That is the filing event.

Company specs: 552B mixture-of-experts (MoE) — 552 billion parameters, with only a slice of the network active per token. Architecture: Causal Encoder–Decoder. Active size: 8B (8 billion) on prefill — the first pass over the prompt — and 16B (16 billion) on decode, when it writes the reply. Context up to 1M tokens — one million tokens, a very long window of text and images. Maximum output 384K (384,000 tokens) on the pricing table.

Pricing for deepseek-flash, per 1 million tokens on the live Models & Pricing page: cache-hit input off-peak $0.003 / peak $0.006; cache-miss input off-peak $0.15 / peak $0.3; output off-peak $0.6 / peak $1.2. A cache hit means the model already has that prompt stored, so input is cheaper. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — the windows when the higher rates apply.

DeepSeek says V4.1-Flash outperforms V4 Pro on its internal multi-party tests for performance, cost, speed, and total time — that is the company’s claim. The news post still described routing deepseek-v4-pro to V4.1-Flash at Flash rates after 12:00 Beijing time on 14 September 2026. The live pricing-page footnote now says that, in response to user demand, DeepSeek will continue the V4 Pro API after 14 September with unchanged billing and will give further notice before any change.

Weights are on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash. Cross-lab “beats GPT…” leaderboard lines stay DeepSeek’s claims, not a desk finding.