Skip to content

models.md

Latest commit

 

History

History
60 lines (38 loc) · 3.41 KB

models.md

File metadata and controls

60 lines (38 loc) · 3.41 KB
Raw

Model Support

ZenMind is designed to work well with modern Chinese model ecosystems while keeping the runtime provider registry flexible.

DeepSeek V4 Agent Configuration

ZenMind uses the current DeepSeek V4 model IDs:

Model key Provider Protocol Model ID Context Max output Pricing source
deepseek-v4-flash deepseek OPENAI deepseek-v4-flash maxInputTokens: 1048576 (1M) maxOutputTokens: 393216 DeepSeek API Docs
deepseek-v4-pro deepseek OPENAI deepseek-v4-pro maxInputTokens: 1048576 (1M) maxOutputTokens: 393216 DeepSeek API Docs

The runtime registry uses modelId, provider, protocol, isReasoner, isFunction, maxInputTokens, and pricing.inputCacheHit/inputCacheMiss/output. These fields are verified by the Agent Platform: maxInputTokens is loaded as the model context window, and the pricing fields are used by usage cost estimation for cache-hit input, cache-miss input, and output tokens.

The DeepSeek provider compatibility layer preserves reasoning_content for DeepSeek-style reasoning messages. ZenMind does not carry a workaround for an already-fixed upstream bug here; it only keeps provider-specific request and response compatibility fields that the runtime reads.

DeepSeek V4 reasoning uses reasoning_effort=max for the highest supported thinking effort. If a caller uses the legacy xhigh level, the compatibility path maps it to max.

Official DeepSeek Pricing

Pricing is verified against DeepSeek API Docs and listed in USD per 1M tokens:

Model Input cache hit Input cache miss Output
deepseek-v4-flash $0.0028 $0.14 $0.28
deepseek-v4-pro $0.003625 $0.435 $0.87

Local registries may also carry CNY billing metadata for runtime cost display, but reviewer-facing pricing in this document follows the official DeepSeek API Docs USD values. Provider secrets, including API keys from local registry files, must not be copied into documentation.

Priority Models

The first product narrative should highlight:

  • DeepSeek V4, using deepseek-v4-pro and deepseek-v4-flash.
  • MiMo, with usage mapping support for provider-specific cache fields.
  • MiniMax M3, as a priority future-facing MiniMax model family.
  • Qwen and Bailian, including coder and general reasoning models.
  • MiniMax office workflows through document, spreadsheet, PDF, and presentation skill environments.

Why It Matters

ZenMind treats model choice as a runtime capability:

  • Agents can carry default model configuration.
  • Chat runs can override model and reasoning effort.
  • The web client can show model options for CODER agents.
  • Usage telemetry carries model key, token details, reasoning tokens, cache details, and estimated cost.

Provider Registry Direction

Model support should stay registry-driven:

  • Providers define base URL, API key handling, protocol behavior, and default model.
  • Models define model key, provider key, model ID, protocol, vision support, reasoning support, context window, and pricing metadata.
  • Compatibility mapping handles provider-specific usage fields so UI usage remains stable.

MiniMax Office Path

ZenMind also treats MiniMax as more than a chat model provider. The local sandbox can expose office-oriented skill paths for PDF, XLSX, DOCX, and PPT workflows, giving agents a runtime designed for document-heavy tasks.