Llama 3.2 lands on Fireworks: the 11B and 90B vision models cost the same as text models, with images billed as 6,400 text tokens
Synopsis
Fireworks announced support for Meta's Llama 3.2 family, covering the text-only 1B and 3B models and the multimodal 11B Vision and 90B Vision models, with inference and fine-tuning available on its platform. That generation was not previously offered there; the 11B and 90B multimodal models are now priced the same as text-only models, and images are charged as text tokens, typically counted as 6,400 text tokens for the Llama 3.2 vision models. Fine-tuning is currently open for the 1B and 3B text-only models, while multimodal fine-tuning is not yet available.
Interpretation
Developers can now call four Llama 3.2 models on Fireworks, two of which read images for captioning, visual question answering and document visual analysis. That generation was not previously on the platform, so text-only and multimodal workloads had to be hosted elsewhere. The stated comparison is throughput on the platform's own serving stack: Llama 3.2 1B at roughly 500 tokens per second and Llama 3.2 3B at 270 tokens per second.
Access is split into three tiers, from pay-per-token serverless calls to private GPUs billed per GPU second to enterprise reserved deployments backed by SLAs. Serverless offers only the instruct variants, while both instruct and non-instruct variants run on on-demand private GPU instances. Meta's Llama Guard 3 models for detecting violating content are also offered on demand alongside these models in the same deployment options.
Calling the multimodal models is brought to the same price as text-only models, with images converted into text tokens rather than billed separately. Image input is handled as a token conversion instead of a separate charge. The conversion depends on image resolution and model, and images for the Llama 3.2 vision models are typically counted as 6,400 text tokens.
Perspective
The report covers Fireworks' hosted support for the Llama 3.2 family, its deployment tiers and billing rules, for developers who need to call or fine-tune these models; fine-tuning for the multimodal models is described as coming soon and is not yet open.
The throughput figures come from the platform's own description of its serving stack and are not accompanied by specific test numbers against vLLM on private GPUs; the token count for an image depends on resolution and model, with 6,400 being the typical value.