Qubit Field Labs trains and serves models for code generation, reasoning, and scientific discovery. Field Harness runs them against your real codebase with full context. Quantum-assisted training is how we push past classical ceilings.
We use quantum computing selectively — for training optimization, loss landscape navigation, and scientific reasoning — not as a general-purpose compute replacement. Every model we serve runs on classical infrastructure.
We train our own architectures with proprietary data pipelines. Each model is optimized for a distinct capability profile and served through our own inference stack.
A general reasoning and code generation model. Trained with process reward models and reinforcement learning from human feedback for reliable multi-step problem solving. Excels at refactoring, debugging, agentic tool use, and structured analysis.
A scientific model for chemistry, materials science, and molecular property prediction. Pretrained on curated scientific corpora with quantum-assisted electronic structure data. Captures physical symmetries that standard training misses.
A coding harness built from the ground up to run our models against real repositories. Not a wrapper around an API — a purpose-built tool with repository awareness, integrated execution, and test feedback loops.
Field Harness indexes your codebase at init — file graph, type hierarchy, call sites. Models receive this context automatically. No manual prompt engineering for repo-level tasks.
Changes are written to disk, then verified against your test suite. Failed tests trigger re-generation with error context. The loop continues until tests pass or a retry limit is reached.
Switch between QF Reasoner 32B for code tasks and QF Matter 14B for scientific workflows. Each model receives task-specific system prompts and tool configurations.
Run Field Harness in your CI pipeline for automated code review, test generation, and migration tasks. Deterministic mode available for reproducible outputs.
pip install field-harness
$ harness init
✓ Repository loaded: 47 files, 12,847 LOC
✓ Model: qf-reasoner-32b (128K context)
✓ Index built: src → tests dependency graph
$ harness run "refactor auth module for OAuth2 PKCE"
Analyzing repository graph…
src/auth/handler.py ──→ src/auth/middleware.py
src/auth/types.py ──→ tests/auth/
Plan generated (4 steps):
1. Create src/auth/oauth2_provider.py
2. Refactor handler.py → PKCE flow
3. Add OAuth2 middleware hook
4. Update test suite
✓ oauth2_provider.py +89 lines
✓ handler.py +52 −31
✓ middleware.py +12
✓ test_handler.py +47
Running: pytest tests/auth/ -q
✓ 18 passed in 0.41sSwitch between models to see how Field Harness adapts its workflow, model selection, and output for different task domains.
Task: Refactor the authentication module to support OAuth2 PKCE flow with token rotation.
$ harness run --model qf-reasoner-32b \
"refactor auth for OAuth2 PKCE with token rotation"
Loading repository context (128K window)…
Relevant files identified via call-graph analysis:
src/auth/handler.py (auth entry point)
src/auth/middleware.py (request validation)
src/auth/types.py (data models)
tests/auth/test_handler.py (24 existing tests)
Generated plan:
1. src/auth/oauth2_provider.py ← new: PKCE + token rotation
2. src/auth/handler.py ← refactor to delegate to provider
3. src/auth/middleware.py ← add token introspection
4. tests/auth/test_handler.py ← 12 new test cases
✓ 4 files modified. +200 −67 lines.
✓ pytest: 36/36 passed (0.52s)
✓ mypy: no new errorsTask: Predict formation energy, band gap, and crystal structure stability for CsPbBr₃ halide perovskite.
$ harness run --model qf-matter-14b \
"predict CsPbBr3 perovskite: formation energy, band gap, stability"
Parsing crystal structure (ABX₃)…
A-site: Cs⁺ | B-site: Pb²⁺ | X-site: Br⁻
Space group: Pm-3m (#221) | Lattice: cubic perovskite
Running quantum-assisted DFT-ML hybrid…
↳ Kohn-Sham SCF: 47 iterations (converged)
↳ Q-enhanced XC correction applied (6-qubit VQE)
Results:
Formation energy: −3.41 eV/atom
Band gap (PBE+Q): 2.26 eV (direct, Γ→Γ)
Lattice constant: 5.87 Å
Thermodynamic stability: Stable (above convex hull by 0.03 eV)
Benchmark vs. experiment:
ΔE_g = +0.06 eV (expt: 2.30 eV)
Wall time: 4.2s (vs. ~23 min pure DFT on same structure)Our training pipeline selectively routes optimization subproblems to quantum processors. Move your pointer to see how quantum signals flow into the model computation graph.
Adjust the rotation angle θ for each qubit and observe how measurement probabilities shift. These are the same state preparations our quantum processors execute during training optimization. The probabilities always sum to 100%.
Measurement in the computational basis (Z) depends only on θ, not φ. State: |ψ⟩ = cos(θ/2)|0⟩ + eiφsin(θ/2)|1⟩. Joint probabilities are products of individual qubit distributions. Our processors prepare these states during variational quantum optimization steps in the training pipeline.
Our research is AI-first. Quantum computing is one of several tools we develop and deploy to improve model training and scientific capabilities.
Using variational quantum eigensolvers and QAOA to compute gradient preconditioning matrices during training. This stabilizes loss landscapes in high-curvature parameter subspaces, reducing training failures and improving convergence for models up to 32B parameters.
Building architectures that encode physical symmetries — conservation laws, gauge invariance, crystallographic constraints — directly into model structure. Our Matter model line predicts molecular properties with quantum-chemical accuracy at a fraction of the compute cost.
Developing training algorithms that selectively offload specific subroutines to quantum processors — attention kernel optimization, sampling from complex distributions, and loss landscape exploration. The goal is targeted quantum advantage, not quantum replacement.
arXiv:2501.08234 · January 2025 · Submitted to ICML 2025
We introduce Q-Grad, a hybrid quantum-classical framework that applies variational quantum circuits to compute gradient preconditioning matrices during LLM pretraining. Applied to models up to 32B parameters, Q-Grad reduces loss spikes by 67% and achieves a 2.4× improvement in effective training throughput compared to standard optimizers with gradient clipping. The method selectively engages quantum processors for high-curvature parameter subspaces identified during training, minimizing quantum compute overhead while maximizing stability gains. We demonstrate consistent improvements across standard language modeling evaluations and show measurable gains on reasoning benchmarks attributable to more stable training dynamics.
From our models and developer platform down through classical accelerators to selectively deployed cryogenic quantum processors. Each layer is purpose-built.
QF Reasoner 32B and QF Matter 14B. Custom architectures, proprietary training recipes, and in-house evaluation suites. Models are trained and served entirely on our infrastructure.
Repository-aware coding platform. CLI, Python SDK, and CI plugins. Manages model routing, context injection, tool execution, and test feedback loops.
Distributed GPU clusters for training and inference. Custom serving stack with optimized attention kernels, KV-cache management, and speculative decoding.
Superconducting qubit systems operating at 15 mK. Used for training optimization subroutines and scientific computation — not as general-purpose compute. Engaged selectively when quantum advantage is analytically expected.
Researchers and engineers with deep experience in large-scale ML, quantum computing, and production systems.
CEO & Co-founder
Former research director at a large-scale ML lab. PhD in computational physics from MIT. Built and led teams training models at frontier scale.
Chief Scientist
15 years in machine learning research. Published 40+ papers on optimization, model scaling, and emergent capabilities. Former faculty at Stanford.
VP Engineering
Built production ML infrastructure serving billions of requests. Expert in distributed systems, inference optimization, and developer tooling. Previously at Stripe.
Head of Quantum Research
Quantum computing pioneer with 20+ patents in superconducting qubit architectures. Former research faculty at Caltech. Leads our quantum-classical integration program.
We're hiring researchers and engineers across ML, quantum computing, and developer tools. All roles are based in our San Francisco office with flexible remote options.
Model Training
Own training pipelines for our next-generation models. Work at the intersection of distributed systems, model architecture, and training stability. You'll integrate quantum-assisted optimization into live training runs.
Quantum Research
Design and implement quantum-classical hybrid algorithms for model training. Focus on variational methods, error mitigation, and practical quantum advantage. Publish your work at top venues.
Developer Tools
Build the next version of Field Harness. Design terminal interfaces, code diff visualizations, and interactive model configuration. React, TypeScript, and a love for developer experience.
Scientific ML
Develop the QF Matter model line. Work with chemists and materials scientists to push AI for science. Experience with molecular dynamics, DFT, or computational chemistry preferred.