Fireworks is the open-model platform for multimodal. Send images through an OpenAI-compatible API, with prompt caching that can cut time to first token (TTFT) by up to 80%. Then fine-tune on your own document types.
You’ll use inference and training for most image work. Orchestration plays a smaller part.
Call vision models on serverless to start, or move to an on-demand deployment for fine-tuned LoRAs; serverless bills per token, and cached input costs less.
When a base model misreads your forms, run vision SFT in the same managed job flow as text SFT and serve the LoRA on an on-demand deployment.
Orchestration barely applies here. If you split image traffic across models, Nexus lets you decide which model answers each request and hold a budget.
Own the vision model you tune on your documents
Healthcare and insurance firms ran 100x cheaper, 1.5x faster than GPT-4o (2024).
AlliumAI runs dozens of fine-tuned LoRA models and pays only for usage (2024).
Fireworks has been Moonshot’s launch partner for Kimi K2.5, K2.6 and K2.7.
Keep the OpenAI SDK and the vision request format you already use.
Ask which vision models are live today and when the model list was last updated.
Your existing OpenAI vision requests should run without a rewrite.
Get per-request image caps and size limits in writing before you design a pipeline.
Ask for token counts by image resolution so you can price a page before you build.
Generic models miss your document taxonomy. Check that you can train on your own images.
Ask for the model, hardware, load and data behind any vision latency number, ours included.
Start with Kimi K2.6, the top pick for vision and document understanding in the Fireworks recommended-models guide.
The request format matches OpenAI’s vision API, so the OpenAI SDK works once you change the base URL and API key.
URLs must be under 5MB and download within 1.5 seconds; base64 must total under 10MB. Use PNG, JPEG, GIF, BMP, TIFF or PPM.
Vision models don’t accept PDFs natively. Render one image per page and send up to 30 images per request.
Use prompt caching, downsize images, use URLs in long chats, put static instructions first and add image metadata.
These image and document models come from the Fireworks recommended-models guide, last updated August 2026. Each one takes image input.
Vision models on Fireworks don’t read PDF files. You render each page as an image and send the pages as image parts, within the per-request cap in step 4. A 12-page contract becomes 12 images. If you need answers across thousands of stored PDFs, that retrieval job belongs on Enterprise RAG.
Every page image bills as input tokens. Most common resolutions run 1,000–2,500 tokens per image; in the docs example, a 1024×1024 image on Qwen2.5 VL is 1,369 tokens. Render pages at the smallest size your model still reads. Multiply tokens per page by the input price in the Models section.
When a base model mislabels your forms, train it on your own pages. Training images must be base64 data URIs. You get a LoRA that runs only on an on-demand deployment. Live merge matches base-model latency and throughput, while multi-LoRA often raises TTFT by 10–30%.
Fireworks takes OpenAI-format vision requests, so your OpenAI SDK code carries over. Add image_url parts to your messages and convert PDFs to page images first.
Fireworks hasn’t published a vision latency benchmark, so we won’t rank ourselves. Test your own images at production load and log TTFT, throughput, model, hardware and date.
Yes. Vision models run on serverless, and the Model Library’s Vision filter lists them. Fine-tuned LoRA models need an on-demand deployment.
Not as files. Convert each page to an image and send the pages as image parts. The guide above covers conversion and token cost.
Images bill as input tokens at your model’s serverless rate. Batch inference costs 50% of serverless pricing.
You choose how each image travels: pass an image URL, or embed it as base64 inline in the request. Open models run with Zero Data Retention by default.
Send images through an OpenAI-compatible API, then fine-tune on your own document types.
