MuskOx · a C library by Auxfield Labs
Follow every possibility. Pay only for the few in play.
MuskOx steps a system forward while keeping track of everything that could be true, and its cost follows the states that are live, not the size of the space. Exact inside a certified range, on a CPU.
Ships as a binary under agreement. Evaluate it free for 45 days on your own hardware.
- Kernel
- Edges along each projection
- Kept
- 12% of samples
- Angles
- 360 backprojected
The idea
Most of the state space is empty at any given moment.
Out of millions of states a system could be in, only a handful are live. MuskOx keeps a short register of just those, steps them forward, and never touches the rest. It isn’t a shortcut: inside a certified range the answer is the one you’d get by tracking everything, and MuskOx computes that range as it runs.
Why it works
The physics stays local.
Half a century of physics says what happens here depends mostly on what’s nearby. MuskOx is built on it.
- 1972
A speed limit
Lieb and Robinson show that in quantum lattices with local interactions, influence spreads at a finite speed.
- 1992
Locality pays
White’s DMRG wins by keeping only the few states that tie one part of a chain to the rest.
- 1996
Nearsightedness
Walter Kohn names it: a local property depends mostly on its surroundings. In 1998 he shares the Nobel Prize in Chemistry, for density-functional theory.
- 2006
Correlations fade
Hastings and Koma, and Nachtergaele and Sims, prove that in gapped ground states, correlations fall off exponentially with distance.
- Now
MuskOx
Keeps its register on the occupied states. When a stretch fills in, a merge step brings it back.
Where it fits
Stop choosing between slow and wrong.
Software that tracks uncertainty usually makes a trade: follow every possibility and run out of compute, or keep a best guess and live with the mistakes. MuskOx follows every possibility and pays only for the ones still in play.
-
Players cross; a robot loses sight of a landmark. MuskOx keeps every “which one is which” alive until the next frame settles it, on the device.
-
Follow a user’s mode or intent beside firmware you’ve already shipped, in a small memory budget and without a GPU.
-
Match the same person across logins, devices and records, and hold “maybe” until the evidence is in instead of merging too early.
-
Ground states and time evolution for chemistry, lattice and spin models, and Markov chains: exact inside the range, on a CPU.
What it isn’t: a replacement for a GPU when nearly every state is in play at once. The advantage comes from the space being mostly empty.
Checked
Exact, against solvers that share none of its code.
Every answer is checked against an independent exact solver. For water, MuskOx needed 133 of its 16,384 possible states, and its answer matches the reference to the ninth decimal place.
Water’s energy in hartree, printed to six decimals. The next three agree too; they part at the tenth, 3×10−10 apart.
All three molecules: H₂, LiH and water, full configuration interaction
| System | Difference | MuskOx (hartree) | PySCF (hartree) | States in play |
|---|---|---|---|---|
| H₂ | 4×10−16 | −1.137276 | −1.137276 | 2 of 16 |
| LiH | 1×10−10 | −7.882395 | −7.882395 | 69 of 4,096 |
| H₂O | 3×10−10 | −75.012578 | −75.012578 | 133 of 16,384 |
Half-filled Hubbard rings
| Ring | U / t | Difference | MuskOx | Exact diagonalization |
|---|---|---|---|---|
| 6 sites | 4 | 1×10−14 | −3.668706179 | −3.668706179 |
| 8 sites | 4 | 1×10−14 | −4.603526300 | −4.603526300 |
| 10 sites | 8 | 3×10−14 | −3.314996729 | −3.314996729 |
Molecules in a minimal basis (STO-3G), against PySCF’s full CI. The suite is a script you can rerun during an evaluation.
In production
Qiviut, our molecular prediction service, folds from scratch on MuskOx, and CAMEO scores its predictions blind every week. See the results ›
In development
Player tracking on broadcast soccer. MuskOx keeps each player’s identity through crossings in about a millisecond a frame on one CPU core, on 39 SoccerNet clips it was never tuned on. An early prototype.
The library
Math you link in, not a service you rent.
MuskOx is a C library. It runs on the CPU you already have, beside the program you already ship. There is no training, no GPU, and nothing hosted.
- Answer
- Exact inside its range. The same answer as tracking every state, inside an operating range the library computes and can report to your program.
- Output
- Deterministic. The same inputs give the same outputs, every run.
- Footprint
- Small. A register of the live states instead of the whole space, so memory follows the problem’s activity, not its size.
- Build
- Plain C11. Links only the C math library and pthreads. No architecture-specific code.
#include <ox.h>
OxConfig cfg = ox_config_default();
OxCtx *ctx = ox_create(&cfg);
/* the states in play, stepped forward */
ox_seed(ctx, kets, n);
ox_evolve(ctx, shells, n_shells, steps, 1);
/* how far from the edge of the range */
OxEnvelope env;
ox_envelope(ctx, &env);
/* the likeliest states */
ox_topk(ctx, 5, top, cap, &n_top);
ox_destroy(ctx);
Access
Try it on a problem you’ve already solved.
The best test is one where you know the answer. Bring it, run MuskOx beside your current method on your own hardware, and compare.
Request access
Tell us about the problem and the method you use today.
Mutual NDA
Signed online, both ways, before anything technical changes hands.
Free evaluation
45 days with the binary, documentation and a license key, on your hardware. No fees.
License
A production license for the binary that fits your deployment.
Run MuskOx on your own problem.
Tell us the problem and the method you use today. An NDA comes first, then 45 days free on your hardware.
Request an evaluationQuestions first
Write to hello@auxfield.com.
Follow along