Models

The compiler writes one BASE_MODEL string. Dense if every weight should be live. MoE if you want active-parameter thrift. LoRA attaches either way.

Live catalog unavailable. The six curated base models below still compile.

Dense — every weight live

Qwen3.5-4BQwen/Qwen3.5-4B · 4B dense · 64KSmallest dense base model before you scale rank or data.For: any jobBuild

Qwen3-8BQwen/Qwen3-8B · 8B dense · 32KDense workhorse when you want every weight live in the loop.For: any jobBuild

Qwen3.5-9BQwen/Qwen3.5-9B · 9B dense + vision · 64KWhen the reward needs screenshots, diagrams, or Lean goals.For: any jobBuild

MoE — sparse-activate

Nemotron-3-Nanonvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 · 30B-A3B · 64KCheap active params, fast tool planning.For: Specialized Tool AgentsBuild

Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B · 35B-A3B + vision · 64KDefault Tinker mid-size: thrifty experts.For: Specialized Tool Agents · Calibrated Forecasting · Formal Reasoning EnginesBuild

DeepSeek-V3.1deepseek-ai/DeepSeek-V3.1 · large MoE · 32KHard jobs: small adapter on a huge base model.For: Calibrated Forecasting · Formal Reasoning EnginesBuild