QF Lab — Research Release 2025

Frontier models,
engineered for precision.

Qubit Field Labs builds advanced AI models and developer tools, leveraging selective quantum-assisted computation to push the boundaries of reasoning, science, and code. Our models run on classical infrastructure; quantum methods augment training and inference where they provide a measurable advantage.

Context window
128K tokens
Coding eval (HumanEval+)
89.2%
Inference throughput
142 tok/s
Training efficiency
2.3× quantum-assisted

Cutting-edge reasoning & scientific models

Our models are trained on classical GPU clusters with selective quantum-assisted optimization for specific scientific and reasoning tasks. Both are available via API and through the Field Harness.

QF Reasoner 32B
32B parameters · 128K context
General reasoning Code generation Tool use Multi-step

Dense transformer with Mixture-of-Experts routing. Trained on 8.5T tokens of code, mathematics, and scientific text. State-of-the-art on coding benchmarks and multi-step reasoning tasks. Ideal for developers, researchers, and engineers.

Context: 128K tokens Eval: 89.2% HumanEval+ Throughput: 142 tok/s Cost: $0.85 / M tokens
QF Matter 14B
14B parameters · 64K context
Chemistry Materials science Quantum-assisted Molecular dynamics

Specialized foundation model for chemistry, drug discovery, and materials design. Pre-trained on 4.2B compounds and 2.1M crystal structures. Quantum-assisted fine-tuning improves conformational sampling and reaction pathway prediction.

Context: 64K tokens Eval: 92.1% QM9 Throughput: 198 tok/s Cost: $0.42 / M tokens

Task: Generate a Python function to compute the DFT of a signal

QF Reasoner 32B produces production-ready code with docstrings, type hints, and unit tests. It leverages its 128K context to reference repository conventions.

$ qf run --model reasoner-32b --task "dft implementation"
$ qf generate --model reasoner-32B --prompt "discrete fourier transform"

import numpy as np
from numpy.typing import NDArray

def dft(x: NDArray[np.complex128]) -> NDArray[np.complex128]:
"""Compute the discrete Fourier transform of a 1D signal."""
N = x.shape[0]
n = np.arange(N)
k = n.reshape((N, 1))
M = np.exp(-2j * np.pi * k * n / N)
return M @ x

# Test
signal = np.array([1, 2, 3, 4], dtype=complex)
print(dft(signal))
✓ Tests passed · 12.3ms inference time

Field Harness — coding, redefined

A terminal-native coding harness optimized for QF models. Select your model, load repository context, execute tools, and get test feedback — all without leaving the command line.

qf-session — zsh
$ qf init --model reasoner-32b
✓ Model loaded: QF Reasoner 32B (128K ctx)
✓ Repository context indexed: 847 files
$ qf generate --task "add rate limiting middleware"
┌─ src/middleware/rate_limit.py
│ from fastapi import Request, HTTPException
│ from time import monotonic
│ ...
│ → 34 lines generated
$ qf test --file src/middleware/rate_limit.py
✓ 5/5 tests passed
• test_limits_exceeded: 12.4ms
• test_window_reset: 8.1ms
$ qf run --tool "lint & typecheck"
✓ mypy: 0 errors · ruff: 0 warnings

Built for real integration

Field Harness connects directly to your repository, indexes your codebase, and gives your team a shared model context. No more copying snippets into a chat UI.

  • Model-aware code generation
  • Automatic test generation & execution
  • Repository-level context indexing
  • Tool execution (lint, typecheck, build)
  • Multi-model switching
  • Quantum-assisted search for complex refactors
pip install field-harness

Quantum-assisted optimization in action

Model computation flows through classical attention layers while quantum-assisted optimization iteratively refines parameter updates. Move your pointer to influence the optimization landscape.

Quantum-assisted optimization landscape ● classical pass

Qubit angle measurement

Adjust two qubit angles and observe the resulting measurement probabilities. This is a simplified model of how quantum-assisted optimization selects candidate parameter updates.

Two-qubit state: |ψ⟩ = cos(θ/2)|00⟩ + e^{iφ} sin(θ/2)|11⟩

Adjust θ (polar) and φ (phase) to see measurement probabilities change.

Measurement probabilities

|00⟩
60.0%
|01⟩
0.0%
|10⟩
0.0%
|11⟩
40.0%
Sum: 100.0%

Simulated sample counts (4096 shots)

|00⟩2458
|01⟩0
|10⟩0
|11⟩1638

Three focused programs

We run targeted research programs where quantum methods can improve AI training, scientific reasoning, and model efficiency. All results are published and open.

Quantum-assisted model optimization

Using variational quantum circuits to identify promising parameter update directions during training, reducing the number of classical gradient steps. Demonstrated 2.3× reduction in training steps for chemical foundation models.

Scientific reasoning

Hybrid models that combine classical transformers with quantum kernels for molecular property prediction, reaction pathway search, and crystal structure generation. Published at NeurIPS 2024 and ICLR 2025.

Hybrid quantum-classical training

Co-training framework where a classical model and a quantum model share representations through a learned projection. Achieves lower loss on scientific benchmarks with 40% fewer parameters.

Selected performance metrics

All figures are fictional and represent internal evaluations on our test suites. They are not claims of real-world performance.

128K tokens
Context length
QF Reasoner 32B
89.2%
HumanEval+ (coding)
QF Reasoner 32B
142 tok/s
Inference throughput
A100 80GB, batch=1
2.3×
Training efficiency
Quantum-assisted vs. pure classical

Featured paper

NeurIPS 2025

Quantum-assisted pre-training for molecular foundation models

E. Vazirani, L. Chen, M. K. Singh, T. Aoki, J. Park

We demonstrate that a hybrid quantum-classical pre-training objective improves downstream performance on 12/14 molecular benchmarks compared to purely classical pre-training. The quantum circuit acts as a regularizer that preserves symmetry information in the latent space. We provide a theoretical analysis of the generalization bound and release all model weights and training code.

Published: December 2025

Full stack, from model to silicon

Our platform spans AI models, developer tools, classical accelerators, and selectively used cryogenic quantum processors.

AI Models

QF Reasoner 32B, QF Matter 14B, and fine-tuned variants

Core

Field Harness

Developer CLI with model-aware code generation and testing

Product

Classical Accelerators

NVIDIA H100 & A100 clusters for training and inference

Workhorse

Cryogenic QPUs

Selective quantum-assisted optimization (16–64 qubits)

Specialized

People behind the lab

Core contributors and scientific advisors.

EV
Elena Vazirani
Chief Scientist, AI
Formerly Google DeepMind
LC
Lei Chen
Head of Engineering
Ex-OpenAI, PyTorch core
MS
M. K. Singh
Quantum Architect
MIT, IBM Quantum
TA
Takahiro Aoki
Research Scientist
RIKEN, Oxford

Open positions

We're looking for people who build at the intersection of AI and quantum.

Research Scientist — AI
AI Research
Develop novel training methods for large language models with quantum-assisted optimization. Strong background in transformer architectures and scientific ML.
Apply →
Quantum Software Engineer
Quantum & Infrastructure
Build the software stack that connects classical ML frameworks with quantum backends. Experience with Qiskit, Cirq, or PennyLane required.
Apply →
Developer Tools Engineer
Product
Design and implement the Field Harness CLI, including model integration, test execution, and repository indexing. Strong Python and systems skills.
Apply →