1---2name: model-reference3description: Use when looking up or comparing current AI model IDs, lifecycle, context/output limits, pricing, modalities, open-weight sizes/licenses, generation settings, deployment memory, or image/video model capabilities across Claude, GPT, Gemini, Qwen, DeepSeek, Llama, Kimi, MiniMax, Gemma, Mistral, GLM, Phi, InternVL, and major media-model families.4---56# Model Reference78Locally maintained, multi-source quick reference for AI model specifications, lifecycle, callable IDs, pricing, deployment constraints, and generation guidance.910This 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.1112## Hard Rules1314- MUST open the relevant file in `references/` before answering model IDs, lifecycle, context/output limits, pricing, sampling parameters, benchmark scores, licenses, or memory requirements.15- MUST state the reference's `Last verified` date when the answer depends on volatile data such as aliases, preview status, pricing, limits, or deprecation.16- MUST distinguish API aliases from pinned snapshots and app/product names from callable API model IDs.17- MUST distinguish native context from provider-extended context and synchronous output limits from batch/long-output modes.18- MUST qualify pricing by provider, unit, prompt tier, cache/batch state, and region when applicable.19- 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.20- MUST distinguish total parameters from active parameters. NEVER estimate MoE weight residency from active parameters alone unless expert offload is an explicit assumption.21- MUST verify official provider docs before production decisions when a reference is older than 30 days or marks a model preview/deprecated.22- NEVER invent model IDs, release dates, context/output limits, prices, knowledge cutoffs, licenses, or generation settings.2324## Fast Lookup2526Run from the directory containing this `SKILL.md`:2728```bash29rg -n "model ID|API ID|alias|snapshot|stable|preview|deprecated|retired|shutdown" references30rg -n "Context|context|Max output|output|Input.*1M|Output.*1M|pricing" references31rg -n "Parameters|active|MoE|VRAM|INT4|FP16|license" references32rg -n "Temperature|TopP|TopK|sampling|generation" references33```3435## Model Families (Open Weights)3637| Family | Creator | File | Focus |38| -------- | -------------- | --------------------------- | ------------------------------ |39| Qwen | Alibaba | `references/qwen.md` | LLM, VLM, MoE, embedding |40| DeepSeek | DeepSeek AI | `references/deepseek.md` | LLM, MoE, reasoning |41| Llama | Meta | `references/llama.md` | LLM, VLM, MoE |42| Kimi | Moonshot AI | `references/kimi.md` | LLM, VLM, MoE, agentic |43| MiniMax | MiniMax AI | `references/minimax.md` | LLM, MoE, coding, agentic |44| Gemma | Google | `references/gemma.md` | LLM, VLM, edge |45| Mistral | Mistral AI | `references/mistral.md` | LLM, audio, coding, reasoning |46| GLM | Zhipu / Z.ai | `references/glm.md` | LLM, VLM, MoE |47| Phi | Microsoft | `references/phi.md` | LLM, VLM, audio, compact |48| InternVL | OpenGVLab | `references/internvl.md` | VLM, vision |4950## Proprietary Models5152| Family | Creator | File |53| ------ | --------- | -------------------------- |54| Claude | Anthropic | `references/claude.md` |55| GPT | OpenAI | `references/gpt.md` |56| Gemini | Google | `references/gemini.md` |5758## Generation Models5960| Category | File |61| -------- | ----------------------------- |62| Image | `references/image-models.md` |63| Video | `references/video-models.md` |6465## Cross-Model Reference6667| Topic | File |68| ---------- | ----------------------------- |69| Overview | `references/overview.md` |7071## Usage Workflow72731. Identify whether the request concerns an API model, downloadable weights, image generation, video generation, or deployment sizing.742. Open the family reference and check `Last verified`, lifecycle, and exact source scope.753. Match the consuming provider, project version, renderer/runtime, region, and requested modality before selecting an ID or configuration.764. Present volatile facts with qualifiers and link the official source.775. Redirect detailed API behavior to the provider-specific documentation skill after selecting the model.
wenerme