AI Models Supported by PandaNpc

Updated

PandaNpc can use built-in chat models supplied by the service and models connected through your own API provider. It can also remotely control coding engines such as Claude Code, OpenAI Codex, and PandaCode. These are related choices, but they are not interchangeable.

This page is a snapshot of the built-in model catalog on August 17, 2026. Model availability changes faster than documentation, so the model selector in the app is authoritative. The public AI model leaderboard is the best place to compare current quality, speed, and price data.

Built-in model catalog

Built-in models can be selected without entering your own provider key. The catalog currently includes:

Provider family Models shown in PandaNpc Good starting point for
OpenAI GPT-5.6 Sol General reasoning, coding, tool-oriented workflows
Anthropic Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 Long-form reasoning, code review, implementation planning
Google Gemini 3.1 Pro Multimodal and long-document work
DeepSeek DeepSeek v4 pro, DeepSeek v4 flash Reasoning or faster cost-sensitive requests
Moonshot AI Kimi Coding k3, Kimi Coding k3 [1m], Kimi K3 Coding and long-context workflows; the [1m] entry explicitly identifies the large-window variant
Zhipu AI GLM-5.2 General-purpose Chinese and coding tasks
Alibaba Qwen 3.7 Max General reasoning, multilingual work, and code

Names in the selector identify a service configuration, not a permanent performance guarantee. Providers may revise underlying model versions, context limits, pricing, rate limits, or regional availability.

Built-in models, API models, and coding engines

Choose the correct layer before comparing names:

Layer Where it runs Authentication and billing Typical use
Built-in PandaNpc model Through the PandaNpc model service Uses your PandaNpc access and applicable service quota Chat, notes, summaries, and supported AI features
Custom-provider model Directly against the Base URL you configure when supported Uses your provider API key and provider billing OpenAI-compatible endpoints, private gateways, Ollama, or LM Studio
Claude Code or Codex engine On your connected computer Uses the account or API configuration of that local CLI Editing repositories, running commands, and agentic coding
PandaCode engine On your connected computer with a configurable backend Uses the model/provider assigned to PandaCode Claude Code-style workflow on DeepSeek, Qwen, private gateways, and other supported APIs

Selecting GPT-5.6 Sol in a chat does not change the model used by a Codex process already running on your machine. Likewise, sharing a Claude Code connection does not grant a separate built-in chat-model quota. See the PandaPaw connection reference for how local coding engines are exposed remotely.

How to choose a model

Start from the task rather than the brand name:

  • Repository-scale coding: use Claude Code, Codex, or PandaCode when the task needs files, shell commands, tests, and tool approvals. Pick the engine's model based on the repository and budget.
  • Complex reasoning or review: begin with a flagship or “pro” model such as GPT-5.6 Sol, Claude Opus 4.8, Claude Sonnet 5, Gemini 3.1 Pro, or DeepSeek v4 pro, then compare actual results on your prompt.
  • Fast iteration: a “flash” configuration such as DeepSeek v4 flash can be a better fit for classification, rewriting, extraction, and repeated small calls.
  • Very long input: choose a catalog entry whose stated context window covers the request, such as Kimi Coding k3 [1m], but do not assume that maximum context delivers maximum accuracy. Retrieval, summaries, and smaller focused inputs are often more reliable and less expensive.
  • Chinese or multilingual work: compare Kimi, DeepSeek, GLM, Qwen, Gemini, Claude, and OpenAI options on representative text. Language quality varies by domain and instruction style.
  • Cost-sensitive workloads: compare input price, output price, cached-input policy, and failure/retry rate. A cheaper token can still cost more if the model needs repeated attempts.

Context window and output limit are different

The context window is the total token budget for instructions, conversation history, tool results, retrieved text, and generated output. The maximum output setting limits only the response. A model advertised with a one-million-token window does not necessarily accept a one-million-token prompt after system instructions and reserved output are included.

For custom PandaCode models, the model form exposes Context Window and Max Tokens. Leave them unset if PandaNpc already knows the model. If you enter custom values, use the provider's documented limits and leave safety room for system prompts and compaction. Setting a window larger than the upstream API supports can make a long-running session fail only after substantial work.

Availability, quotas, and privacy

Before relying on a model in production, confirm:

  1. It appears in your current model selector and is authorized for your account.
  2. Its input types match your task; a text model may not accept images or files.
  3. Its provider quota, rate limit, and billing policy are acceptable.
  4. Your data-handling requirements permit sending the selected content to that provider.
  5. Tool calling and structured output work with the exact model version you selected.

PandaNpc may transmit chat or session content to the provider that powers the selected feature. Review the Privacy Policy and your provider's own retention and training terms before sending confidential material.

Add a model that is not listed

If an OpenAI-compatible or supported provider model is missing, add a provider and then register its exact model identifier. The display label can be human-friendly, but Model Value must match the identifier expected by the upstream API. Follow Configure a custom AI provider and API key for the complete setup and verification flow.

When a model disappears or fails unexpectedly, first check the live selector and provider status. Catalog changes, retired model IDs, expired keys, account limits, and incompatible context settings are more common causes than a PandaNpc client installation problem.