Update: DeepSeek V4.1 Flash is multimodal, live on 12 API providers
Vision was trained in from pre-training, the KV cache is ~890 bytes/token, and on Monday DeepSeek starts routing V4-Pro API traffic straight to Flash.
Copy markdown
Vision was there all along
It takes text and image input natively — DeepSeek says the visual and text embeddings were trained jointly from the start of pre-training, not bolted on after. It's the lab's first non-experimental model with native image input, so the 'cheap coder' is also a cheap 1M-token vision model.
Now on a dozen providers, not just DeepSeek's API
OpenRouter lists it across 12 providers (Fireworks, DeepInfra, Novita, SiliconFlow, Modal, Parasail and more) at $0.15/M in and $0.60/M out off-peak. Baseten added it Sep 11 with the full 1M-token context and NVIDIA Dynamo KV-cache-aware routing — so you can pick a host on latency and price instead of being locked to one endpoint.
Heads-up: V4-Pro reroutes to Flash on Monday
Starting Sept 14, DeepSeek routes v4-pro API requests to V4.1-Flash at Flash pricing until a v4.1-pro ships. If your app pins deepseek-v4-pro, its outputs and your bill shift Monday with no change on your end — pin an explicit version now if you need the current behavior.
The architecture reveal: encoder-decoder plus a lookup table
Reportedly it's 763B total — a ~748B backbone plus a ~197B 'engram' lookup component and vision modules — but only ~8B active on prefill and ~16B on decode. A novel encoder-decoder design compresses the KV cache to ~890 bytes/token, which is how a 1M-token window stays this cheap to serve.
Run it yourself this weekend — but watch the token count
On a 4-GPU box you can reportedly hit 200-300+ tokens/sec at full precision by offloading the engram table to NVMe. The catch: it's verbose, burning ~89k tokens per task versus 55-71k for rivals — still about $0.27/task, but size your output budgets accordingly.