Open weights · Local inference
aipoker.dev is a curated index of Hugging Face checkpoints and Ollama-ready pulls for no-limit Texas hold'em—built around PokerBench (AAAI 2025) and community fine-tunes through 2026.
Listed models
12
Install paths
Ollama · HF · GGUF
UI languages
EN · KO · ZH · JA
These models output actions from text game states. They are research tools—not real-money bots. Follow platform terms and applicable law.
Modern poker LLMs are not monolithic "poker engines." They are instruction-tuned language models that read a structured scenario—position, stacks, board, action history—and emit the next action (fold, check/call, raise with sizing). Quality depends on training data (PokerBench, GTO Wizard traces, self-play preferences) and whether the prompt format matches what the model saw during fine-tuning.
This page tracks publicly hosted weights you can run today: merged SFT checkpoints, LoRA adapters (PEFT), GRPO/DPO refinements, and GGUF builds for Ollama. Each card includes copy-ready commands; GGUF files route through a safe redirect to Hugging Face `resolve/main`.
RZ412/PokerBench ships ~60k preflop and ~500k postflop training rows plus ~11k test scenarios for six-max NLHE. Paper code: pokerllm/pokerbench.
Install Ollama · run `ollama pull` on an hf.co model · point playerpokerbot or your agent at the tag · keep temperature low (~0.1) for stable actions.
Install torch + transformers (+ peft for LoRA adapters) · `huggingface-cli download` for large shards · load with `device_map="auto"` on CUDA/Metal.
Download `.gguf` from a quant repo · write a Modelfile with low temperature · `ollama create` · run interactive tests with paper-format prompts.
ollama pull hf.co/YiPz/Qwen3-4B-pokerbench-sft:latest
huggingface-cli download YiPz/llama3-8b-pokerbench-sft --local-dir ./llama3-pokerbench
Public Hugging Face repos tagged or trained on PokerBench / GTO distillations. Filter by how you plan to run inference.
ollama pull hf.co/YiPz/Qwen3-4B-pokerbench-sft:latest
Python load
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "YiPz/Qwen3-4B-pokerbench-sft" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
Engines and apps that consume these weights—useful after you download a model.
Official PokerBench codebase from the AAAI 2025 paper—dataset loaders, fine-tuning recipes, and evaluation harness for NLHE decision accuracy.
Play up to six-max against local Ollama models with PokerKit rules, web UI, and PHH-compatible action parsing—pairs naturally with YiPz Qwen3 PokerBench SFT.
Event-driven Texas hold'em engine for building LLM poker servers—OpenAI, Anthropic, or OpenRouter agents with pot-odds tooling.
Local-first HU NLHE research stack with CFR solvers, ML policy heads, FastAPI backend, and React cockpit—not an LLM weight but complements model benchmarking.
