Model catalog
Every model Mortar serves, with the context window a request can use and sample prices per million tokens. Choose a cache hit rate to see what prompt caching does to the input price.
6 models
Sample pricing in US dollars per million tokens, not a quote. The cached input price assumes 50% of prompt tokens are read from the prefix cache at 0.1× the input price, and the rest are processed and written to it at 1.25×.
| Model | Family | Context | Input | Output | Input at 50% cached | Status |
|---|---|---|---|---|---|---|
Qwen 2.5 7B InstructQwen/Qwen2.5-7B-Instruct |
Alibaba Qwen | 4KModel supports 32K | $0.0313 | $0.2373 | $0.0211 | |
Llama 3.1 8B Instructmeta-llama/Llama-3.1-8B-Instruct |
Meta Llama | 128K | $0.0347 | $0.2654 | $0.0234 | |
Mistral 7B Instructmistralai/Mistral-7B-Instruct-v0.3 |
Mistral AI | 32K | $0.0297 | $0.2278 | $0.0200 | |
Qwen 2.5 32B InstructQwen/Qwen2.5-32B-Instruct |
Alibaba Qwen | 32K | $0.1529 | $1.1046 | $0.1032 | |
Llama 3.3 70B Instructmeta-llama/Llama-3.3-70B-Instruct |
Meta Llama | 128K | $0.6253 | $4.5671 | $0.4221 | |
DeepSeek-V3deepseek-ai/DeepSeek-V3 |
DeepSeek | 128K | $1.1057 | $7.7809 | $0.7463 |
How these prices are worked out
- Each price is the hourly list price of the GPUs a model is sized for, divided by the tokens they process in an hour at 65% utilization.
- Input costs less than output because processing a prompt takes far less GPU time per token than generating a reply.
- A prompt token read from the prefix cache costs 0.1× the input price, and every token that has to be processed is written to the cache for the next request at 1.25×. Below a hit rate of about 22%, caching costs more than it saves. Cache-friendly prompts raise your hit rate.
- Context is the window a request can use on this deployment, as GET /v1/models reports it.