qwen3-1.7b-chess-coach-mlx

Qwen3 chess coach

A level-calibrated, engine-grounded chess coach fine-tuned to turn engine truth into intuitive, tier-appropriate teaching.

What it is

This model is a local-first AI chess coach designed to provide rating-calibrated move review. Given a position, a student's rating tier (beginner, intermediate, or advanced), the move played, and full-strength engine analysis, the coach recommends one sound teaching move and explains it in plain language. It avoids centipawns, engine jargon, and GM-only lines, focusing entirely on what a human at the specified tier can understand and apply.

How it was built

The training pipeline relies on three main components to ensure reliability.

The latest iteration uses Qwen3-32B as the base model, fine-tuned via QLoRA. Earlier versions use Qwen3-1.7B merged and quantized to 4-bit MLX.

Tier-appropriate move behavior

The core strength of this model is its ability to differentiate move selection by rating tier. Instead of blindly recommending the engine's absolute best move, the coach steers beginners toward human-findable moves (guided by Maia likelihoods) and reserves the sharpest, most complex engine lines for advanced players.

Evaluation

The project relies on a definitive, zero-leakage benchmark scored by a cross-family council of frontier models (GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro) alongside deterministic objective metrics.

Because evaluation numbers are actively regenerating through a grand eval, please refer to the live artifacts for the most current results (as of July 8, 2026):

Usage

For the MLX 1.7B version:

from mlx_lm import load, generate

model, tok = load("khoilamalphaai/qwen3-1.7b-chess-coach-mlx")
messages = [
    {"role": "system", "content": "You are a chess coach doing move review."},
    {"role": "user", "content": "Student tier: beginner. [Engine-grounded prompt details.]"},
]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tok, prompt=prompt, max_tokens=512))

This repository also bundles the prior v3 32B QLoRA adapter in the v3-lora-qwen3-32b/ subfolder: load the Qwen/Qwen3-32B base in 4-bit and attach it with PEFT.

The shipped v4 32B QLoRA adapter (the strongest tier-appropriate move-selection checkpoint) is published as its own repository: chess-coach-32b-v4-qlora.

Links

Limitations