LLM Providers

💡TL;DR

Notemd supports 36 LLM providers across 4 transport types. All API calls use raw HTTP (no SDKs) with automatic retry, transport fallback, and 5-minute response caching. Per-task model selection lets you use cheap models for concept extraction and powerful models for research. No API key needed for local providers (Ollama, LMStudio).

This is part of the Obsidian AI Knowledge Management Guide.

Provider Categories

Cloud Providers

Provider	Default Base URL	Default Model	Temperature
DeepSeek	api.deepseek.com	deepseek-v4-pro	0.5
OpenAI	api.openai.com/v1	gpt-4o	0.5
Anthropic	api.anthropic.com	claude-3-5-sonnet-20240620	0.5
Google	generativelanguage.googleapis.com/v1	gemini-2.0-flash-exp	0.5
Mistral	api.mistral.ai/v1	mistral-large-latest	0.5
xAI	api.x.ai/v1	grok-4	0.7
Cerebras	api.cerebras.ai/v1	gpt-oss-120b	0.5
Fireworks	api.fireworks.ai/inference/v1	kimi-k2p5	0.5
Nebius	api.studio.nebius.com/v1	openai/gpt-oss-120b	0.3

Gateway / Proxy Providers

Provider	Default Base URL	Default Model	Notes
OpenRouter	openrouter.ai/api/v1	anthropic/claude-3.7-sonnet	Adds HTTP-Referer + X-Title headers
GitHub Models	models.github.ai/inference	gpt-4o-mini	Adds X-GitHub-Api-Version header
AIHubMix	aihubmix.com/v1	gpt-4o-mini	Adds APP-Code header
PPIO	api.ppinfra.com/v3/openai	qwen/qwen3-32b
Hugging Face	router.huggingface.co/v1	openai/gpt-oss-120b
Vercel AI Gateway	ai-gateway.vercel.sh/v1	anthropic/claude-sonnet-4.5
Requesty	router.requesty.ai/v1	anthropic/claude-3-7-sonnet-latest
Together	api.together.xyz/v1	meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo
LiteLLM	localhost:4000/v1	your-proxy-model	Optional API key
New API	localhost:3000/v1	gpt-4.1	Optional API key
OpenAI Compatible	(user-configured)	your-model-id	Custom endpoint

China Providers

Provider	Default Base URL	Default Model	Notes
Qwen	dashscope.aliyuncs.com/compatible-mode/v1	qwen3-235b-a22b	Alibaba DashScope
Qwen Code	dashscope.aliyuncs.com/compatible-mode/v1	qwen3-coder-plus	Coding-focused
Doubao	ark.cn-beijing.volces.com/api/v3	(endpoint ID required)	ByteDance Volcengine
Moonshot	api.moonshot.cn/v1	kimi-k2-0905-preview	Kimi
Xiaomi MiMo	api.xiaomimimo.com/v1	mimo-v2.5-pro
GLM	open.bigmodel.cn/api/paas/v4	glm-5	Zhipu, China endpoint
Z AI	api.z.ai/api/paas/v4	glm-5	Zhipu, international endpoint
MiniMax	api.minimaxi.com/v1	MiniMax-M2.7
Baidu Qianfan	qianfan.baidubce.com/v2	ernie-4.5-turbo-32k
SiliconFlow	api.siliconflow.cn/v1	Qwen/QwQ-32B	Model hosting platform
Huawei Cloud MaaS	api.modelarts-maas.com/v1	DeepSeek-V3

Doubao requires an Ark endpoint ID (format: ep-xxxxxxxx-xxxx-xxxx). The default placeholder must be replaced.

Local Providers

Provider	Default Base URL	Default Model	Notes
Ollama	localhost:11434/api	llama3	No API key. NDJSON streaming. Model discovery via /api/tags
LMStudio	localhost:1234/v1	local-model	Sends "EMPTY" as bearer token. OpenAI-compatible discovery
OVMS	localhost:8000/v3	openvino-model	OpenVINO Model Server
Azure OpenAI	(user-configured)	gpt-4o	Deployment name in URL. api-key header

Per-Task Model Selection

Notemd supports independent provider and model for each task. Enable with useMultiModelSettings: true.

Available Tasks

Task	Settings Key	Default Provider
Add Links (wiki-link insertion)	addLinksProvider / addLinksModel	DeepSeek
Research (web search + summarize)	researchProvider / researchModel	DeepSeek
Generate from Title	generateTitleProvider / generateTitleModel	DeepSeek
Translation	translateProvider / translateModel	DeepSeek
Diagram Generation	summarizeToMermaidProvider / Model	DeepSeek
Concept Extraction	extractConceptsProvider / Model	DeepSeek
Original Text Extraction	extractOriginalTextProvider / Model	DeepSeek

Resolution Order

1. If useMultiModelSettings is on → use the task-specific provider + model
2. If task model override is empty → fall back to the provider's default model
3. If task provider is invalid → fall back to the global activeProvider
4. If useMultiModelSettings is off → all tasks use the global activeProvider

Recommended Configurations

# Researcher: power for research, cheap for everything else
activeProvider: DeepSeek
extractConceptsProvider: DeepSeek
researchProvider: OpenAI        # GPT-4o for better research quality
summarizeToMermaidProvider: DeepSeek

# Privacy-first: local only
activeProvider: Ollama
addLinksProvider: Ollama
extractConceptsProvider: Ollama

# Budget bilingual
activeProvider: DeepSeek
translateProvider: Google      # Gemini Flash for translation
researchProvider: SiliconFlow  # Cheaper China-based routing

API Call Architecture

Transport Layers

1. Obsidian requestUrl — primary (built-in, works on all platforms)
2. Desktop HTTP — Node.js http/https modules (streaming + long requests)
3. Web fetch — fallback when Node.js modules unavailable

The plugin automatically cascades between transports on transient network errors.

Retry Logic

• Normal mode (enableStableApiCall: false): Single attempt. On transient failure, auto-escalates to stable retry.
• Stable retry mode (enableStableApiCall: true): Up to 4 attempts (1 + 3 retries) with 5s intervals.
• Transient errors that trigger retry: err_connection_closed, err_connection_reset, err_timed_out, econnreset, econnaborted, etimedout, socket hang up, network request failed
• Non-retryable: HTTP 400, 401, 403, 404

Response Caching

5-minute in-memory LRU cache keyed by provider + model + temperature + prompt + content. Subsequent identical requests within 5 minutes return cached results instantly.

Reasoning Model Handling

Models that don't support the system role (DeepSeek R1, OpenAI o1/o3/o4/gpt-5) automatically have system + user prompts combined into a single user message.

DeepSeek-specific: thinkingEnabled adds { thinking: { type: 'enabled' } } to the request. reasoningEffort (low/medium/high) is passed as reasoning_effort. Validates that max output tokens ≥ 8000 when thinking is on.

Token Estimation

estimateTokens(text): Math.ceil(text.length / 4) — 4-chars-per-token heuristic. Used to enforce maxResearchContentTokens before sending to LLM.

No cost tracking is implemented.

Model Discovery

Mode	Providers	Endpoint
ollama-tags	Ollama	/api/tags
google-models	Google	Models list API
anthropic-models	Anthropic	Models list API
openai-compatible-models	DeepSeek, Qwen, Mistral, etc.	/v1/models
openrouter-models	OpenRouter	Model list API
github-models	GitHub Models	Inference API
none	Azure OpenAI	Disabled (requires deployment name)

Quick Start

1. Pick a provider — DeepSeek (China), OpenAI (International), or Ollama (Local/Free)
2. Enter API key — Settings → Notemd → Provider → API Key
3. Test connection — Click the test button
4. Optional: Per-task models — Enable useMultiModelSettings and configure individually

For Ollama: no API key needed. Just make sure ollama serve is running.

---

Next Steps

• 📝 Concept Notes — Uses per-task extraction model
• 🎨 Diagrams — Uses per-task diagram model
• 🔍 Research — Uses research-specific model
• ⚙️ Configuration — Full settings reference