Projects | Juan Cruz-Benito

Qiskit Code Assistant

Tue, 01 Apr 2025 00:00:00 +0000

A family of LLMs (8B–24B parameters) specialized for quantum code generation. Models built on IBM Granite, Mistral, and Qwen foundations, fine-tuned on curated Qiskit datasets including Python scripts, Jupyter notebooks, and synthetic Q&A pairs.

Open-source model family:

Mistral Small 24B — Largest model, highest accuracy
Qwen2.5 Coder 14B — Strong coding foundation
Granite 3.x 8B — Efficient, multiple versions
GGUF quantized versions available for local deployment

Achieves 46.53% on Qiskit HumanEval—significantly outperforming competing models (24.75%–39.6%). Supports natural language to code (“define a Bell circuit and run it on ibm_brisbane”) and intelligent autocomplete.

We created and open-sourced Qiskit HumanEval and Qiskit HumanEval Hard—benchmarks with 150+ tasks each, now used industry-wide for evaluating quantum code LLMs. Integrated into VS Code and JupyterLab.

qiskit-ibm-transpiler

Sat, 01 Mar 2025 00:00:00 +0000

AI-powered quantum circuit optimization that outperforms traditional heuristics. Uses reinforcement learning to achieve near-optimal synthesis of Linear Function, Clifford, and Permutation circuits—orders of magnitude faster than SAT solvers. Our Pauli Network synthesis delivers over 2× reduction in two-qubit gate count, with average improvements of 20% and up to 60% on the Benchpress benchmark.

Supports hardware-aware routing up to 133 qubits and works as a drop-in replacement for standard Qiskit transpilation. Available as both local execution (with our open-source RL models) and cloud-based services.

Achievements:

(full-stack platforms)
(after just 1 year public)
• 56 releases • Apache 2.0

Qiskit Gym

Sat, 01 Feb 2025 00:00:00 +0000

Gymnasium-compatible RL environments for training AI agents to synthesize quantum circuits. The framework that powers the AI transpiler passes achieving state-of-the-art results in qiskit-ibm-transpiler.

Three synthesis environments:

Permutation Synthesis — Minimal SWAP gate implementations respecting hardware coupling
Linear Function Synthesis — CNOT-optimal decomposition of Boolean linear functions
Clifford Synthesis — Hardware-efficient implementations of Clifford group elements

Hardware-aware design matches real quantum device coupling maps. High-performance Rust backend enables fast training. Supports PPO, AlphaZero, and custom policies with built-in TensorBoard visualization.

The agents trained with this framework achieve near-optimal synthesis up to 65 qubits—orders of magnitude faster than SAT solvers.

Qiskit MCP Servers

Wed, 01 Jan 2025 00:00:00 +0000

Production-ready Model Context Protocol servers that give AI assistants direct access to quantum computing infrastructure. Built on FastMCP with async-first architecture, full type safety, and 65%+ test coverage.

Four specialized servers:

Qiskit Server — Circuit creation, manipulation, and transpilation
Code Assistant Server — AI-assisted quantum programming via Granite models
IBM Runtime Server — Hardware access, job execution, and backend selection
AI Transpiler Server — RL-powered circuit optimization

Enables AI systems to autonomously generate quantum code, select optimal backends, execute circuits on real hardware, and provide expert quantum computing guidance—all through a standardized protocol.