Pick the right, reproducible foundation model for your agent experiment.
Turn paper goals, resource constraints, and priorities into a shareable research task; walk a verifiable path across candidate models, evidence sources, and the compare tray.
Choose by your reproduction goal
Keep the original paper model and environment
Use the same model ID and checkpoint as the paper; control environment variables and only replace non-model factors such as dataset version or random seed.
Swap models but control key variables
Validate whether the method still holds on an equivalent or newer-generation model; record how model differences affect results.
Re-validate the method on a current-generation model
Rerun an older method on today’s open-weight or API flagship to see if the conclusion degrades or improves across generations.
Start from a scenario
LoRA on a single 24GB GPU
Weight update = LoRA, single 24GB card, open weights only.
Pre-fill workspace →Multi-GPU RL or self-evolution training
Weight update = RL, multi-GPU, open weights preferred.
Pre-fill workspace →Find an open-weight coding model
Open weights only, suited for code / tool-use experiments.
Pre-fill workspace →Reproduce a Qwen2.5 paper
Strict reproduction path, Chinese capability prioritized.
Pre-fill workspace →Replace an old model with a current generation
Modern rerun, focused on current-generation models.
Pre-fill workspace →Compare API and local models
Low cost first, comparing API and local candidates.
Pre-fill workspace →Recent records by release date
- claude-2Official catalog record marks Claude 2 as retired.Data event
- Gemini 3.6 FlashGoogle · GeminiReleased
- Kimi K3Moonshot AI · KimiReleased
- GPT-5.6 LunaOpenAI · GPTReleased
- GPT-5.6 SolOpenAI · GPTReleased
- GPT-5.6 TerraOpenAI · GPTReleased
- Grok 4.5xAI · GrokReleased
Flagship family changes
Qwen
Qwen3.6Qwen3.6 is the latest Qwen generation, with Qwen3.6-35B-A3B released as the open-weight flagship and served via API as qwen3.6-flash.
Llama
Llama 4Llama 4 is the latest Llama generation, with Scout and Maverick as the initial open-weight models.
DeepSeek
DeepSeek-V4DeepSeek-V4 is the latest DeepSeek generation, with Pro and Flash variants released as open weights.
Kimi
Kimi K3Kimi K3 is the latest Kimi open-weight model recorded in the official catalog
Model-space overview
Browse all models by release date, inference hardware, parameter scale, and evidence status. Read the supply layer first, then layer on paper adoption.