nvidia: built-in model catalog is stale — many currently-available NVIDIA API models are missing, and custom providers are blocked on v2 servers #51031

Problem

The built-in nvidia provider model list (sourced from models.dev) lags noticeably behind NVIDIA's own API catalog (https://build.nvidia.com/models). A number of models that are currently available on NVIDIA cannot be selected in /models. At the same time, on servers detected as v2 protocol, the app's Custom provider dialog is blocked by Custom providers are unavailable on this server, so users cannot fill the gap with their own configuration. Together this makes some usable NVIDIA models completely unselectable.

Reproduction

1. Connect to a server detected as v2 protocol (hosted instance or opencode serve).
2. Open /models, go to the nvidia provider and search for any of the IDs below → none exist.
3. (Optional) Try adding these models via the Custom provider dialog → saving fails with Custom providers are unavailable on this server.

Missing models (still listed on NVIDIA API catalog as of 2026-09)

Evidence: https://build.nvidia.com/models

Some historical model IDs were also renamed, so configs using the old IDs break, e.g. nvidia/ai-synthetic-video-detector (catalog now uses nvidia/nvidia/synthetic-video-detector).

Impact on users

Example: in one user's opencode.json, all ~57 models under provider.nvidia.models are absent from the built-in catalog on shared/hosted servers (verified one by one against /models and the model-list endpoint). Because v2 servers do not allow adding custom providers from the app UI, there is no in-UI workaround; users must spin up their own local server instead.

Expected behavior

1. Refresh the nvidia model list in a timely manner: sync NVIDIA's API catalog on a fixed cadence (or bump the models.dev data source), covering at least the models in the table above.
2. Relax the v2-server restriction: allow adding/overriding custom models for built-in providers on v2-protocol servers; or at least show a hint in /models when configured models are not displayed, explaining why (e.g. custom providers are unavailable on this server).

Related code locations

• packages/app/src/components/dialog-custom-provider.tsx:134 — saving a custom provider throws when protocol !== "v1"
• packages/app/src/utils/server-protocol.ts — v1/v2 protocol detection
• The nvidia built-in model list is generated from models.dev data (nvidia/-prefixed entries in /models)

Environment

• OpenCode version: v2 (app connects to the server with protocol = "v2")
• Where observed: /models model picker and the Custom provider dialog

中文摘要(Chinese summary)

内置 nvidia 提供商的模型列表(来源 models.dev)滞后于 NVIDIA 自身 API catalog(https://build.nvidia.com/models),
/models 里选不到一批当前仍可用的 NVIDIA 模型;而 v2 协议服务器上,"自定义提供商"对话框会被
Custom providers are unavailable on this server(此服务器上无法使用自定义提供商)拦截,用户无法自行补齐缺失模型。

• 复现:连接 v2 协议服务器 → /models → nvidia 提供商搜索上述模型 ID(全部不存在);用"自定义提供商"添加时保存报错。
• 缺失示例(2026-09 仍见于 build.nvidia.com):riva-translate-4b-instruct-v2(catalog 仅有已退役 v1.1)、
  nemotron-parse-2.0、nemotron-3-embed-1b、nemotron-3.5-content-safety、ising-calibration-1.5-31b、
  diffusiongemma-26b-a4b-it、wan2.2-animate-2-14b、parakeet-tdt-0.6b、nemotron-3-ultra-550b-a55b、
  cosmos3-nano(-reasoner)、qwen-image(-edit-nvpcb-ovsl2sl)、nemotron-ocr-v2、chatterbox-multilingual-tts、
  deepseek-ai/deepseek-v4.1-flash 等。
• 影响:本项目 opencode.json 中 provider.nvidia.models 的约 57 个模型在共享/托管服务器上全部不出现;
  v2 服务器又禁自定义提供商 → 页面内无补丁手段,只能自己起本地服务器。
• 期望:① 按固定节奏同步 NVIDIA catalog(或升级 models.dev 数据源),覆盖上表模型;
  ② 放宽 v2 服务器限制,允许为内置提供商新增/覆盖自定义模型,或在 /models 提示"配置了但未显示"的原因。
• 代码位置:packages/app/src/components/dialog-custom-provider.tsx:134(protocol !== "v1" 即抛错)、
  packages/app/src/utils/server-protocol.ts(v1/v2 判定)。