Research Lab · Est. 2023

Frontier models,
quantum-accelerated.

We train and serve state-of-the-art language models for reasoning and science. Quantum computing accelerates key bottlenecks in training, inference, and molecular simulation — giving our models a structural advantage without replacing the classical stack that does the heavy lifting.

Trained on classical accelerators · Optimized with quantum · Served at scale

128K
Model Context Length
Fictional benchmark
94.2%
Coding Eval Score
HumanEval+ (fictional)
340 tok/s
Inference Throughput
H100 batch-8 (fictional)
18%
Quantum-Assisted Training Efficiency
vs. classical-only baseline

Built for reasoning, tuned for science.

Two models designed for distinct domains. Both trained with quantum-assisted optimization on our internal cluster.

QF-REASONER-32B

QF Reasoner 32B

A general-purpose reasoning and coding model. Excels at multi-step logic, code generation and debugging, mathematical proof, and structured tool use. Ideal as a backbone for developer-facing applications.

Parameters
32B
Context
128K tokens
License
QF Research
Precision
BF16 / INT4
Try Reasoner
QF-MATTER-14B

QF Matter 14B

A specialist model for computational chemistry and materials science. Trained on curated scientific corpora and fine-tuned with quantum simulation feedback for molecular property prediction and reaction pathway analysis.

Parameters
14B
Context
32K tokens
License
QF Research
Precision
BF16
Try Matter

See each model in action.

Select a model to view a realistic Field Harness session with that model's strengths.

Select Model

QF Reasoner 32B
General reasoning and coding
QF Matter 14B
Chemistry & materials science

From tokens to qubits and back.

Classical neural computation identifies optimization targets; quantum sampling explores the solution landscape. Move your cursor across the visualization to see the interplay.

Model gradient flow → quantum-enhanced optimization → updated weights

Tune two qubits. Watch the probabilities.

Adjust θ and φ to rotate each qubit on the Bloch sphere. The measurement probabilities update in real time and sum to 100%.

|0⟩ |1⟩ Qubit 0 |0⟩ |1⟩ Qubit 1
|00⟩
50.0%
|01⟩
0.0%
|10⟩
50.0%
|11⟩
0.0%
Sum of probabilities: 100.00%

Three directions. One lab.

We focus on where quantum computation measurably improves AI capabilities.

Program 01

Quantum-Assisted Model Optimization

Using variational quantum circuits and quantum approximate optimization (QAOA) to navigate high-dimensional loss landscapes. Targets weight initialization, learning rate schedules, and architecture search — not replacing gradient descent, but seeding it with better starting points.

Program 02

Scientific Reasoning

Training models that can formulate hypotheses, plan experiments, and interpret results in chemistry and materials science. Integrates classical simulation data with quantum chemistry benchmarks for grounded, verifiable scientific outputs.

Program 03

Hybrid Quantum-Classical Training

Co-designing classical and quantum components that execute in the same training loop. Focus on reducing communication overhead, handling qubit decoherence gracefully, and ensuring graceful degradation when quantum resources are unavailable.

Field Harness.

A coding harness optimized for QF models. Understands repository context, executes tools, manages multi-file edits, and streams structured feedback — all running against our inference endpoints or your own deployment.

$ pip install qf-harness

📂 Repository Context

Automatically indexes your codebase, dependency graph, and test structure. Passes only relevant context to the model, reducing token waste and improving edit quality.

🔧 Tool Execution

Runs linters, formatters, tests, and custom scripts within a sandboxed environment. Model sees real output and iterates — no more guessing at compiler errors.

🔄 Diff-Based Editing

Generates minimal, reviewable diffs rather than rewriting entire files. Integrates with git for atomic commits, branch-aware context, and conflict detection.

NeurIPS 2024 — Spotlight

Hierarchical Training with Quantum Approximate Optimization for Language Model Pre-training

A. Chen, R. Matsuda, L. Okonkwo, P. Vanderstraeten, D. Wei — Qubit Field Labs & Stanford
December 2024

We present a method for integrating quantum approximate optimization (QAOA) into the pre-training of large language models at three specific stages: weight initialization via quantum sampling of low-energy configurations in random energy landscapes, learning rate warmup shaped by quantum circuit evaluation of loss curvature, and architecture search over attention head configurations. Across 7B–32B parameter models, we observe 12–23% reductions in total training compute to reach matched downstream benchmarks, with the largest gains on reasoning-intensive tasks.

TRAINING LOSS CURVES Classical QAOA Loss Steps (×10³)

Full stack. Selectively quantum.

Quantum processors are one layer in the stack — used where they help, bypassed where they don't.

🧠

QF Models & Field Harness

Reasoner 32B, Matter 14B — and the developer platform that makes them usable.

Inference & Serving Layer

vLLM-based serving with speculative decoding, continuous batching, and model-parallel sharding.

Classical Accelerator Cluster

H100 / A100 nodes for training and inference. Handles the vast majority of computation.

Cryogenic Quantum Processors

Selectively invoked for optimization subroutines, molecular simulation, and architecture search.

Our quantum processors operate at 15 mK in dilution refrigerators and are accessed via a cloud API. They are used for approximately 8–12% of total training compute — targeted at stages where quantum sampling provides measurable advantage. When quantum resources are unavailable, the system degrades gracefully to classical-only baselines.

The people behind the models.

AC

Ada Chen

Head of Research

Previously led ML systems at a national lab. PhD in quantum information from MIT. Focused on making quantum optimization practical for large-scale training.

RM

Ryo Matsuda

Principal Engineer

Systems architect with deep experience in GPU clusters and distributed training. Builds the infrastructure that connects classical and quantum compute.

LO

Lisa Okonkwo

Product Lead, Field Harness

Former core contributor at a major code intelligence platform. Designs the developer experience that makes QF models accessible in real workflows.

Build what's next.

We're hiring across research, engineering, and product. Remote-friendly with offices in San Francisco and Tokyo.

Quantum ML Engineer

Research · Full-time

Design and optimize quantum-classical training loops. You'll work at the intersection of variational quantum circuits and large-scale model training.

Apply

Systems Engineer, Field Harness

Engineering · Full-time

Build the tool execution runtime, repository indexing pipeline, and multi-model orchestration layer for our developer platform.

Apply

ML Research Scientist

Research · Full-time

Advance scientific reasoning capabilities in QF Matter. Develop training recipes, evaluation suites, and benchmark datasets for chemistry and materials science.

Apply