Janus-Pro

Janus-Pro (Multimodal)

Checked on 15 Nov 2026 against DeepSeek API docs.

Janus-Pro is an autoregressive framework that unifies multimodal understanding and generation by decoupling visual encoding into separate pathways while using one transformer stack. The Hugging Face model card describes SigLIP-L for 384×384 understanding inputs and a discrete VQ tokenizer (16× downsample) for generation. This directory does not list benchmark scores, API token prices, or context length for Janus-Pro because those figures were not in that card.

API list price
No per-token API list price for Janus-Pro on the current DeepSeek API pricing page.
Context
Not stated on the Janus-Pro-7B model card
Parameters
1.5B and 7B variants; active Dense (see Janus-Pro-1B / Janus-Pro-7B weights)
Architecture
SigLIP-L understanding + discrete VQ generation; DeepSeek-LLM backbone

No per-token API list price for Janus-Pro on the current DeepSeek API pricing page.

Running Janus-Pro locally

Janus-Pro is a research multimodal model family (1.5B and 7B checkpoints on Hugging Face) built on DeepSeek-LLM backbones. The Janus-Pro-7B model card explains the decoupled design: SigLIP-L encodes 384×384 images for understanding, while a discrete VQ tokenizer (16× spatial downsampling, LlamaGen family) handles image generation inside one autoregressive stack.

Quick-start commands and inference scripts live in the Janus GitHub repositorylinked from the card. This directory does not list Janus-Pro on the current DeepSeek API pricing table; treat it as open weights for experimentation, not as a hosted deepseek-* API model name.

For API-accessible multimodal coding workloads today, DeepSeek documents deepseek-flash (V4.1-Flash) with vision inputs — see the V4.1-Flash model page and official pricing. Janus-Pro remains useful when you need unified text+image generation research outside the commercial API surface.

All models · API pricing · Off-peak pricing guide · LLM API price comparison · Model names · FAQ