Inference API — Sequences

The slow brain

Nothing here is wrapped. No system prompt is injected, no message is rewritten, no reply is parsed. Messages go in, one message comes back — in the same envelope the action endpoint answers with.

plan.pyplain http

import requests

plan = requests.post(
    "https://api.generalsequences.com/v1/chat",
    headers={"Authorization": f"Bearer {SEQUENCES_KEY}"},
    json={
        "model": "accounts/sequences/models/claude-fable-5",
        "messages": [{"role": "user", "content": [
            {"type": "text", "text": "What should the arm do next?"},
            {"type": "image_url", "image_url": {"url": frame}},
        ]}],
    },
).json()["message"]["content"]

The fast brain — observation in, action chunk out

One request shape for every policy. A feed-forward VLA reads the latest frame per camera; a world-action model reads a window of them. You send the frames you have and the response tells you what was used.

act.pyno equivalent elsewhere

chunk = requests.post(f"{BASE}/act", headers=HEAD, json={
    "model": "accounts/sequences/models/pi05-droid",
    "observation": {
        "images": [
            {"view": "exterior_1", "data": frame_a, "timestamp_ms": 1000},
            {"view": "wrist_left",  "data": frame_b, "timestamp_ms": 1000},
        ],
        "proprioception": {"joint_positions": q, "gripper": [0.8]},
        "instruction": plan,
    },
}).json()["action_chunk"]

# 15 steps × 32 dof, covers 1.00 s at 15 Hz, joint_absolute
# send the next request before covers_seconds elapses or the arm starves

The same body serves world models. Repeat a view with increasing timestamps and it becomes a video. Isaac reads three; the world models read one and imagine the rest — Cosmos3 Edge seventeen frames ahead, DreamZero thirty-three. Send fewer than a model reads and the earliest frame is repeated to fill the window — and the response says so in warnings, because padding that happens quietly is the hardest kind of wrong to find.

What one GPU carries

Weight and frames read, published the same way for every model. Vendors measure latency on whatever hardware they own, so their figures compare machines as much as models.

Model	Family	Action	Context	Parameters
pi05-droid	VLA	32d × 15 @ 15 Hz	1 frame	4.14 B
lingbot-vla-v2-6b	VLA	55d × 25 @ 30 Hz	1 frame	6.38 B
xiaomi-robotics-1-5b	VLA	75d × 20 @ 25 Hz	1 frame	5 B
isaac-0-5	VLA	32d × 20 @ 25 Hz	3 frames	36 B
groot-n1-7-3b	VLA	32d × 16 @ 30 Hz	1 frame	3 B
cosmos3-edge-policy-droid	WAM	8d × 16 @ 15 Hz	1 frame, 17 imagined	3.86 B
dreamzero-agibot	WAM	32d × 48 @ 30 Hz	1 frame, 33 imagined	14 B

What we deliberately do not do

No retry loop, no failure detector, no sub-goal scheduler. Those depend on things only you know — whether the task counts as finished, whether the scene can be reset, how many attempts are worth paying for. Anthropic measured a supervisor placed at the action layer performing worse than the policy alone, because it overrode the policy more often than was warranted. We route, meter, and translate. The loop is yours.