models

Guide for choosing models in Tinker — available model families, model types (base, instruction, reasoning, hybrid, vision), architecture (dense vs MoE), and how to match renderers to models. Use when

SKILL.md

GitHub

Model Selection

Help the user choose the right model for their task.

Reference

  • docs/model-lineup.mdx — Full model listing with types, sizes, and architecture
  • tinker_cookbook/model_info.py — Model metadata and renderer mapping

Available models

Qwen family

Model Type Arch Size
Qwen/Qwen3.5-397B-A17B Hybrid + Vision MoE Large
Qwen/Qwen3.5-35B-A3B Hybrid + Vision MoE Medium
Qwen/Qwen3.5-27B Hybrid + Vision Dense Medium
Qwen/Qwen3.5-4B Hybrid + Vision Dense Compact
Qwen/Qwen3-235B-A22B-Instruct-2507 Instruction MoE Large
Qwen/Qwen3-30B-A3B-Instruct-2507 Instruction MoE Medium
Qwen/Qwen3-30B-A3B Hybrid MoE Medium
Qwen/Qwen3-30B-A3B-Base Base MoE Medium
Qwen/Qwen3-32B Hybrid Dense Medium
Qwen/Qwen3-8B Hybrid Dense Small
Qwen/Qwen3-8B-Base Base Dense Small
Qwen/Qwen3-4B-Instruct-2507 Instruction Dense Compact
Qwen/Qwen3-VL-235B-A22B-Instruct Vision MoE Large
Qwen/Qwen3-VL-30B-A3B-Instruct Vision MoE Medium

Llama family

Model Type Arch Size
meta-llama/Llama-3.3-70B-Instruct Instruction Dense Large
meta-llama/Llama-3.1-70B Base Dense Large
meta-llama/Llama-3.1-8B Base Dense Small
meta-llama/Llama-3.1-8B-Instruct Instruction Dense Small
meta-llama/Llama-3.2-3B Base Dense Compact
meta-llama/Llama-3.2-1B Base Dense Compact

Nemotron family

Model Type Arch Size
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 Hybrid MoE Large
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 Hybrid MoE Medium

Other families

Model Type Arch Size
openai/gpt-oss-120b Reasoning MoE Medium
openai/gpt-oss-20b Reasoning MoE Small
deepseek-ai/DeepSeek-V3.1 Hybrid MoE Large
deepseek-ai/DeepSeek-V3.1-Base Base MoE Large
moonshotai/Kimi-K2-Thinking Reasoning MoE Large
moonshotai/Kimi-K2.5 Reasoning + Vision MoE Large

How to choose

By task type

  • Instruction tuning / chat SFT: Start with an Instruction model (e.g., Llama-3.1-8B-Instruct, Qwen3-30B-A3B-Instruct-2507)
  • RL with verifiable rewards (GRPO): Use Instruction or Hybrid models — they already follow instructions
  • Reasoning / chain-of-thought: Use Reasoning or Hybrid models (Kimi-K2-Thinking, Qwen3-8B)
  • Full post-training pipeline: Start with a Base model (e.g., Qwen3-8B-Base, Llama-3.1-8B)
  • Vision tasks: Use Vision or Hybrid+Vision models (Qwen3.5-35B-A3B, Qwen3-VL-*)
  • Distillation (student): Use a Base model as student
  • Quick prototyping: Use compact models (Llama-3.2-1B, Qwen3.5-4B)

By cost

Prefer MoE models — they're much more cost-effective than dense models because cost scales with active parameters, not total parameters. For example, Qwen3-30B-A3B (MoE, 3B active) is cheaper than Qwen3-32B (Dense, 32B active) despite similar quality.

Model types explained

  • Base: Pre-trained on raw text. For research or full post-training pipelines.
  • Instruction: Fine-tuned for instruction following. Fast inference, no chain-of-thought.
  • Reasoning: Always uses chain-of-thought before visible output.
  • Hybrid: Can operate in both thinking and non-thinking modes.
  • Vision: Processes images alongside text. See /renderers skill for vision input handling.

Size categories

  • Compact: 1B–4B parameters
  • Small: 8B parameters
  • Medium: 27B–32B parameters
  • Large: 70B+ parameters

Renderer matching

Every model needs a matching renderer. Always use the automatic lookup:

from tinker_cookbook import model_info

renderer_name = model_info.get_recommended_renderer_name(model_name)

Never hardcode renderer names — the mapping is maintained in model_info.py.

Learning rate by model

Use hyperparam_utils.get_lr(model_name) for model-specific LR recommendations. See the /hyperparams skill for details.

Common pitfalls

  • MoE models are cheaper than dense — prefer them unless you have a specific reason
  • Base models need full post-training (SFT + alignment) to be useful for chat
  • Instruction models are best for tasks where you want to start from a capable baseline
  • Vision models require ImageChunk in messages — see /renderers skill
  • Llama models require HF_TOKEN for tokenizer download (gated on HuggingFace)