server/catalog.py

"""The models this harness offers, and what is true about each one.

Kept as data rather than fetched at boot: the Space must come up and be usable
when huggingface.co is slow or down, and a model list that sometimes appears is
worse than one that is occasionally a day stale. `insight.py` fetches the live
provider mapping on demand, for the panel that exists to show exactly that.
"""

from __future__ import annotations

from dataclasses import asdict, dataclass, field


@dataclass(frozen=True)
class Model:
    id: str
    label: str
    vendor: str
    providers: list[str] = field(default_factory=list)
    reasoning: bool = False
    context_tokens: int | None = None
    note: str = ""

    def to_json(self) -> dict:
        d = asdict(self)
        d["contextTokens"] = d.pop("context_tokens")
        return d


# Verified live on the HF router on 2026-08-28. `providers` is the routing
# order reported by the hub, not a preference of ours.
MODELS: tuple[Model, ...] = (
    Model(
        id="zai-org/GLM-5.3",
        label="GLM-5.3",
        vendor="Z.ai",
        providers=["zai-org", "baseten"],
        reasoning=True,
        context_tokens=200_000,
        note="Released 2026-08-28. Emits a separate reasoning stream and "
             "defaults to maximum thinking budget.",
    ),
    Model(
        id="zai-org/GLM-5.3-Flash",
        label="GLM-5.3 Flash",
        vendor="Z.ai",
        providers=["novita", "together", "fireworks-ai", "zai-org", "baseten"],
        reasoning=True,
        context_tokens=200_000,
        note="The small sibling. Useful for checking whether a slow answer is "
             "the model or the route.",
    ),
    Model(
        id="deepseek-ai/DeepSeek-V4-Pro-0813",
        label="DeepSeek V4 Pro",
        vendor="DeepSeek",
        providers=["novita", "together", "fireworks-ai", "baseten", "deepinfra"],
        reasoning=True,
        context_tokens=163_840,
        note="The comparison model: the latest DeepSeek as of 2026-08-13.",
    ),
    Model(
        id="deepseek-ai/DeepSeek-V4-Flash-0731",
        label="DeepSeek V4 Flash",
        vendor="DeepSeek",
        providers=["novita", "together", "fireworks-ai", "baseten", "deepinfra"],
        reasoning=True,
        context_tokens=163_840,
        note="Fast DeepSeek, for a like-for-like against GLM-5.3 Flash.",
    ),
)

DEFAULT_MODEL = "zai-org/GLM-5.3"

_BY_ID = {m.id: m for m in MODELS}


def get(model_id: str) -> Model | None:
    return _BY_ID.get(model_id)


def is_allowed(model_id: str) -> bool:
    """Model ids arrive from the browser and are pasted into an upstream URL."""
    return model_id in _BY_ID