Pay only for what you
train and run

Say goodbye to idle GPU bills and per-seat pricing. Foundry meters every fine-tune and prediction in credits — so you only pay for the compute your models actually use.

# Fine-tune & run models

Bring a leading open model or run a hosted one. Fine-tune it into your own named destination, then pay credits per prediction when you run it. Every cover image and video clip below was generated on Replicate.

Llama 3.1 8B cover
meta logo

meta / Llama 3.1 8B

Instruction-tuned all-rounder to fine-tune for chat, agents, and tool use.

60 crfine-tune2 cr/ prediction
Mistral 7B cover
mistralai logo

mistralai / Mistral 7B

Fast, efficient base with great price-to-quality for high-volume assistants.

55 crfine-tune2 cr/ prediction
Gemma 2 9B cover
google logo

google / Gemma 2 9B

Lightweight open base with strong reasoning for tutors and support agents.

65 crfine-tune2 cr/ prediction
Zephyr 7B β cover
huggingface logo

huggingface / Zephyr 7B β

Aligned, helpful chat base. Fine-tune on your voice and ship in minutes.

55 crfine-tune2 cr/ prediction
Qwen2.5 7B cover
ollama logo

ollama / Qwen2.5 7B

Multilingual base with strong coding and math. Great general-purpose tune.

55 crfine-tune2 cr/ prediction
Claude 3.5 Haiku coverRun only
anthropic logo

anthropic / Claude 3.5 Haiku

Frontier quality for hard reasoning and long context. Run-only.

3 cr/ prediction
DeepSeek R1 coverRun only
deepseek logo

deepseek / DeepSeek R1

Open reasoning model with chain-of-thought. Run-only, metered per call.

3 cr/ prediction
Gemini 1.5 Flash coverRun only
google logo

google / Gemini 1.5 Flash

Very fast multimodal model with a huge context window. Run-only.

2 cr/ prediction
Sonar Large coverRun only
perplexity logo

perplexity / Sonar Large

Search-grounded answers with live citations. Run-only, per prediction.

2 cr/ prediction
Video

lightricks / LTX Video

Fast text-to-video. Generate short clips from a single prompt.

8 cr/ clip
Video

minimax / video-01

Cinematic text-to-video with smooth motion and high detail.

40 cr/ clip

# Training compute

Fine-tuning cost scales with the base model. Larger bases cost more credits and take longer, but reason better. Re-training an existing model is cheaper than the first run.

Base modelParamsFirst fine-tuneRe-trainTypical time
Foundry Mini 1.5Bfoundry/mini-1.5b1.5B87 credits$3.9263 credits$2.84~3–8 min
Foundry Instruct 3Bfoundry/instruct-3b3B117 credits$5.2784 credits$3.78~6–15 min
Foundry Lingual 3Bfoundry/lingual-3b3B122 credits$5.4988 credits$3.96~6–15 min

# Usage rates

Everything you do maps to a simple credit cost. One credit ≈ $0.05, based on the Builder pack.

ActionAPICredits≈ USD
First fine-tunetrainings.create55 credits$2.48
Re-fine-tunetrainings.create (existing)40 credits$1.80
Singer voice traintrainings.singer85 credits$3.83
Singer re-traintrainings.singer (refit)65 credits$2.93
Chat / predictionrun()1 credit$0.05
Workspace actionworkspace.*2 credits$0.09
Song generationsongs.generate12 credits$0.54

# Credit packs

Top up any time. Credits never expire, and unused credits roll over. Start free with welcome credits — no card required.

Starter

$9one-time

200 credits

$0.045 / credit

  • 3 first fine-tunes
  • 5 re-trains
  • 200 predictions
Get Starter

Studio

$69one-time

2,000 credits

$0.035 / credit

  • 36 first fine-tunes
  • 50 re-trains
  • 2,000 predictions
Get Studio