We are partnering with MiniMax to bring MiniMax M2.7 to Lilac.
Fast inference, no markup.
Frontier-speed inference on idle enterprise GPUs — same output quality, lower cost.
inference.py
from openai import OpenAI
client = OpenAI(
base_url="htt",
api_key="sk_...",
)
response = client.chat.completions.create(
model="openai/gpt-5.4",
messages=[{"role": "user","content": "Hello!"}],
)
Real-time model performance.
Recent performance snapshot across every endpoint we serve.
Model status
| Model | TPS | TTF (P50) | Availability |
|---|---|---|---|
| MiniMax M2.7 (minimaxai/minimax-m2.7) | 158 tok/s | 0.17 s | OK Operational |
| Kimi K2.6 (moonshotai/kimi-k2.6) | 120 tok/s | 0.33 s | OK Operational |
| GLM 5.1 (zai-org/glm-5.1) | 102 tok/s | 0.43 s | OK Operational |
| Gemma 4 (31B) (google/gemma-4-31b-it) | 58 tok/s | 0.41 s | OK Operational |
Updated every 30 seconds from recent API traffic.
Pricing
Pay per token. No commitments.
We route inference to idle enterprise GPUs — hardware that's already powered on and paid for. No reserved capacity markup. You only pay for what you use.
| Model | Context | Input | Output |
|---|---|---|---|
| MiniMax M2.7 | FP8 | 200K | $0.30/M |
| Kimi K2.6 | INT4 | 262K | $0.70/M |
| GLM 5.1 | FP8 | 203K | $0.90/M |
| Gemma 4 (31B) | BF16 | 262K | $0.11/M |
More models coming soon. The economics
Why we're cheaper.
Most GPU clusters run at 30–50% utilization. Lilac routes your inference to that idle capacity — hardware that's already powered on and paid for.
- Idle GPUs are already powered on — no cold start overhead.
- Providers set competitive rates to monetize spare capacity.
- Lilac routes to the best available GPU automatically.
- You only pay per token. No reserved instances, no commitments.
Start running inference in minutes.
No contracts, no commitments. Swap your base URL and pay less for the same output quality.