Model Reference
Locally maintained, multi-source quick reference for AI model specifications, lifecycle, callable IDs, pricing, deployment constraints, and generation guidance.
This skill is not synchronized from a single provider. Each reference is a dated manual snapshot built from provider-primary documentation, official repositories/model cards, release announcements, and technical reports.
Hard Rules
- MUST open the relevant file in
references/ before answering model IDs, lifecycle, context/output limits, pricing, sampling parameters, benchmark scores, licenses, or memory requirements.
- MUST state the reference's
Last verified date when the answer depends on volatile data such as aliases, preview status, pricing, limits, or deprecation.
- MUST distinguish API aliases from pinned snapshots and app/product names from callable API model IDs.
- MUST distinguish native context from provider-extended context and synchronous output limits from batch/long-output modes.
- MUST qualify pricing by provider, unit, prompt tier, cache/batch state, and region when applicable.
- MUST label benchmarks as vendor-reported unless an independent primary evaluation source and setup are cited. Do not treat different benchmark settings as directly comparable.
- MUST distinguish total parameters from active parameters. NEVER estimate MoE weight residency from active parameters alone unless expert offload is an explicit assumption.
- MUST verify official provider docs before production decisions when a reference is older than 30 days or marks a model preview/deprecated.
- NEVER invent model IDs, release dates, context/output limits, prices, knowledge cutoffs, licenses, or generation settings.
Fast Lookup
Run from the directory containing this SKILL.md:
rg -n "model ID|API ID|alias|snapshot|stable|preview|deprecated|retired|shutdown" references
rg -n "Context|context|Max output|output|Input.*1M|Output.*1M|pricing" references
rg -n "Parameters|active|MoE|VRAM|INT4|FP16|license" references
rg -n "Temperature|TopP|TopK|sampling|generation" references
Model Families (Open Weights)
| Family |
Creator |
File |
Focus |
| Qwen |
Alibaba |
references/qwen.md |
LLM, VLM, MoE, embedding |
| DeepSeek |
DeepSeek AI |
references/deepseek.md |
LLM, MoE, reasoning |
| Llama |
Meta |
references/llama.md |
LLM, VLM, MoE |
| Kimi |
Moonshot AI |
references/kimi.md |
LLM, VLM, MoE, agentic |
| MiniMax |
MiniMax AI |
references/minimax.md |
LLM, MoE, coding, agentic |
| Gemma |
Google |
references/gemma.md |
LLM, VLM, edge |
| Mistral |
Mistral AI |
references/mistral.md |
LLM, audio, coding, reasoning |
| GLM |
Zhipu / Z.ai |
references/glm.md |
LLM, VLM, MoE |
| Phi |
Microsoft |
references/phi.md |
LLM, VLM, audio, compact |
| InternVL |
OpenGVLab |
references/internvl.md |
VLM, vision |
Proprietary Models
| Family |
Creator |
File |
| Claude |
Anthropic |
references/claude.md |
| GPT |
OpenAI |
references/gpt.md |
| Gemini |
Google |
references/gemini.md |
Generation Models
| Category |
File |
| Image |
references/image-models.md |
| Video |
references/video-models.md |
Cross-Model Reference
| Topic |
File |
| Overview |
references/overview.md |
Usage Workflow
- Identify whether the request concerns an API model, downloadable weights, image generation, video generation, or deployment sizing.
- Open the family reference and check
Last verified, lifecycle, and exact source scope.
- Match the consuming provider, project version, renderer/runtime, region, and requested modality before selecting an ID or configuration.
- Present volatile facts with qualifiers and link the official source.
- Redirect detailed API behavior to the provider-specific documentation skill after selecting the model.