MarsCompute
Self-hosted model inference on heterogeneous accelerators. Workload-specific pilot capacity and production routing to be agreed.
MarsCompute sells inference on Liquid Inference. A buyer names one of its models as marscompute/<model>, or names the model and pays the lowest offer that meets the buyer's rules.
Facts
- Handle
- marscompute
- Website
- https://marscompute.ai/
- Terms of service
- https://marscompute.ai/terms
- Privacy policy
- https://marscompute.ai/privacy
- On the exchange since
- 2026-10-30
- Models
- 6
- Offers
- 6
- Regions
- global, us
- Precisions (declared)
- fp8, provider-managed
- Lowest price on
- 2 of 6 models
Reliability, last 3 days
- Probes answered
- 100.00%
- Hours with an answer
- 100.0%
- Median response time
- 2.7s
Checked by the exchange
1 of 6 offers match the fingerprint of the model. 0 have a finished evaluation.
The exchange sends fixed prompts to each offer and compares the answers with a baseline for the model. It also runs benchmarks against the offers and keeps the results.
Data handling
- Zero retention, declared by the provider
- No training, declared by the provider
Models
- GLM 5.3: declared precision provider-managed, global, $1.400 input and $4.400 output per million tokens, $1.400 cache read per million tokens, 1M tokens of context
- GLM 5.3 Flash: declared precision provider-managed, global, $0.150 input and $0.500 output per million tokens, $0.150 cache read per million tokens, 1M tokens of context
- Kimi K3: declared precision provider-managed, global, $3.000 input and $15.000 output per million tokens, $3.000 cache read per million tokens, 1M tokens of context
- Mimo V2.6 Flash: declared precision provider-managed, global, $0.105 input and $0.210 output per million tokens, $0.105 cache read per million tokens, 1M tokens of context
- MiniMax M3.1 Flash: declared precision provider-managed, global, $0.150 input and $0.600 output per million tokens, $0.150 cache read per million tokens, 1M tokens of context
- Qwen: Sea Lion V4.5 27B It: declared precision fp8, us, $0.600 input and $3.600 output per million tokens, $0.600 cache read per million tokens, 8,192 tokens of context