Speed decides how much each user is worth.
Agents are token-hungry and multi-step. At typical speeds users watch them type, do less per session, and churn. Faster inference makes every user more valuable.
500+ tps.
Inference for agents.
Frontier models at high throughput and lower cost. Built for background agents that work across long horizons.
MODELS FOR AGENTS
BUILT FOR LONG HORIZONS
Run more agent work on the same hardware.
Agents reason, call tools, retry, and continue. Amtal lowers the cost of the full task without asking them to use a weaker model.
Capacity decides how many you can serve.
The GPU crunch is real, and there is no new supply of NVIDIA GPUs. So growth gets capped: you waitlist users your competitors also can't serve.
Trade latency for lower token prices.
Migrate latency tolerant workloads towards different response patterns for dramatically lower token costs.
1
From Silicon
Amtal optimizes and tunes kernels for the exact shapes that production inference workloads require.
to every task.
2
Verify every sharding.
For each machine, Amtal enumerates parallelism and placement choices, removes configurations that do not fit, prices their collectives, and ranks the viable routes.
3
Warm the cache. Beat the burst.
Amtal learns burst patterns from your production traffic. When a burst is detected, it copies the shared prefix KV cache across workers over RDMA, then routes requests to the least-loaded warm worker.
Shared context. More warm workers. Less waiting.
Warm the fleet before the burst.
Production traffic → Learned trigger → RDMA → Least load
Adaptive trigger
Replicate when the expected time saved exceeds the copy time.
Resident KV · Capacity availableLeast-loaded warm worker
Once the copies are ready, spread the burst across workers sharing the same prefix.
MODEL CATALOG
Frontier models.
Hot when they ship.
Open models selected for coding, reasoning, and long-horizon agent workloads.
DeepSeek
V4 Flash 0731
Fast, efficient inference for coding, reasoning, and long-horizon agent workloads.
Contact us for enterprise capacity and dedicated deployments
Contact salesDeepSeek
V4.1 Flash
Multimodal Mixture-of-Experts with a 1M-token context window for long-horizon agent workloads.
Contact us for enterprise capacity and dedicated deployments
Contact salesStart the task.
Let agents finish.
Submit long-running work with the OpenAI client you already use. Background responses return immediately while Amtal runs the task asynchronously.
- Background task execution
- Pollable response status
- Retry-safe submissions
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.amtalml.com/v1",
api_key=os.environ["AMTAL_API_KEY"],
)
response = client.responses.create(
model="deepseek-v4-flash-0731",
input="Run the test suite, diagnose failures, and propose fixes.",
background=True,
metadata={"completion_window": "balanced"},
extra_headers={"Idempotency-Key": "agent-task-123"},
)
print(response.id, response.status)Get the best inference
for the shape of your workload.
Bring your longest-running agent workload. Keep its intelligence and lower its inference cost.