Steinborn IT GmbH

Steinborn IT GmbH · Acquisition and investment enquiries

We built it. Now it is for sale.

I am selling because I cannot fund the next stage of this research. The German innovation funders I approached declined to support it. I first contacted SPRIND on 15 June 2026. Today is 24 October, more than 18 weeks later, and the outcome is a refusal. I am putting the work in front of people who can buy it and take it further.

Computing technology with patents pending. Language models on CPUs. Models of brain areas and individual cells. New mathematics and simulated universes. Bid on one of 13 parts, several parts or the whole company. You can also propose an investment. Fourteen preprints are public. The measurements are below. The unpublished methods go to the buyer.

Deadline
Tuesday, 1 December 2026 at 23:59 CET
Choose what you want
One of 13 parts, a combination or the company. Investment proposals are welcome.
Selection
The highest qualifying total, under the terms below. Every deal is subject to contract.
Ownership
Exclusive rights to the work we own, with the stated licence exceptions. We stop work on each part sold.
Company warranty
Steinborn IT GmbH warrants our measurements as stated, on the terms agreed in the contract.
Make an offer
sales@steinborn.it

The numbers

What one server did

Our benchmark machine has one AMD EPYC 9455 processor, 128 GB of RAM and no GPU. The individual entries give the workloads and limits. Figures attributed to model vendors come from their own tests on different machines.

26.8 billioncandidate evaluations each secondSoftware Evolver. Short synthetic programs on one CPU, using the earlier build included in the sale.
1,904,711cities searched in 60 secondsWorld TSP. Best run 2.22 % above the best known tour on 6 cores.
0.098 %above a proven optimum85,900 cities. About 7 minutes from scratch on 48 cores.
150,000simulated players within a 50 ms cycleA 16 × 16 km world. This run did not include the network stream.
37 tokens/swith a model of 27 billion parametersTernary Bonsai 2 27B. Approximate measured rate on 42 CPU cores, without a GPU.
61,549model universes containing light and matterNo faster than light behaviour found in the tested cases.
14public research preprintsPreprints in mathematics, physics, chemistry and neuroscience.
3pending patent applicationsSFA, Software Evolver and LLM Analyzer.

The research has outlasted the money

I built this work through Steinborn IT and directed the research across computing, language, neuroscience, mathematics and physics. There are working systems here, published proofs and unfinished lines that deserve to be finished.

The reason for this sale is funding. I do not have the money to carry the research further. My applications to German innovation funders did not secure it. SPRIND declined after a process that began on 15 June 2026. By today, 24 October, more than 18 weeks have passed.

I find that outcome hard to square with the ambition to fund frontier research. My standard for advanced work is what it can demonstrate. The papers and measurements on this page are what I put forward. You can judge them for yourself.

I am done waiting for permission to continue. The work is now available to whoever makes the strongest qualifying offer, as individual parts, a package or the company. An investor can also propose funding its continuation. Offers close on Tuesday, 1 December 2026 at 23:59 CET.

You buy the work at its present stage. Steinborn IT GmbH will warrant the accuracy of its reported measurements under the terms of the sale contract. The unfinished work is identified, and its future belongs to whoever takes it on. If you want me to tackle your own research instead, we can discuss that separately.

Pascal Eric SteinbornFounder of Steinborn IT GmbH

The black box

The idea is part of what you are buying

No disclosure in exchange for an NDA

Once you have seen how an invention works, you cannot unsee it. A confidentiality agreement gives legal remedies. It does not restore a secret. We will not disclose unpublished code, concepts or formulas before a completed sale, even under NDA.

Put a price on the work

There is no technical data room and no extended inspection of the methods before you bid. Read the papers, consider the measurements and make an offer. Ownership, existing licences and the legal terms still have to be settled in the contract.

The uncertainty is yours to price

You decide what the work is worth to you on these terms. Fourteen papers are available to read. Three patent applications are pending. Steinborn IT GmbH warrants its own measurements as stated, with the scope, remedies and time limit agreed in the contract. It does not promise what unfinished research will become. Our promise to keep the unpublished work confidential after the sale is part of what you buy.

A demonstration, if you must

If you need to see a result before bidding, we can agree on a neutral demonstration under NDA for a part that can be shown safely. We run it live on our server. You see the agreed inputs, outputs and measurements. You receive no code, formulas, internal traces, executable files or access to the server. The demonstration terms must protect against recording, extraction and attempts to reconstruct the methods.

Some parts, including the brain model, cannot be demonstrated without giving away the work. Those are sold blind. A demonstration gives you no reservation or priority. Your offer competes with every other offer. For equal offers on otherwise equivalent terms, the bidder who buys blind takes priority.

Computing technology

The foundations and the systems built with them

SFA, the Software Evolver and the LLM Analyzer are available for licensing and have pending patent applications. The applications are included with the relevant acquisitions. Patent pending means an application has been filed, not that a patent has been granted. Proof of filing can be supplied on request with technical details withheld.

Part 1

SFA

Patent pending

Computing across platforms with exact, deterministic results. It stores data without loss and recovers the original data exactly.

Available for licensing now.

Part 2

Software Evolver

Patent pending

Searches for compact programs, formulas and decision rules that solve a given task. It runs on server CPUs without a language model.

  • 26.8 billion evaluations per second on one AMD EPYC 9455 using 96 threads. This measured short synthetic programs. One evaluation means one candidate program tested on one test case. The result comes from an earlier engine build, which is included in the sale.
  • Used in our mathematics research. The published work includes a formula verified as an exact symbolic identity.

Available for licensing now and used in our own research.

Part 3

LLM Analyzer

Patent pending

Examines the stored weights of an AI model without prompting it, generating text or running inference.

Verification and mapping work on public Gemma, Qwen and IBM Granite models. The full exposure audit remains unfinished.

Part 4

Route optimization

Working software

Builds valid routes through very large lists of stops. Every returned route is checked independently for validity. The figures below use standard travelling salesman benchmarks without additional constraints such as delivery windows or vehicle capacities.

  • 1,904,711 cities in the World TSP. A search of 60 seconds on 6 cores found a tour 2.22 % above the best known tour. That was the best run. A repeat reached 2.26 %.
  • 0.228 % and 0.342 % above the proven optima for 10,639 and 24,978 cities after 10 seconds on 6 cores. At 60 seconds, the gaps were 0.122 % and 0.182 %.
  • Under one second from starting the program to writing the result on 6 cores. The gaps above the proven optima were 0.61 % for 10,639 cities and 0.94 % for 24,978 cities.
  • Eight simultaneous jobs on one server, each using 6 cores. After 10 seconds, all eight routes through 24,978 cities were within 0.40 % of the proven optimum.
  • 0.098 % above the proven optimum for 85,900 cities, starting from scratch and running for about 7 minutes on 48 cores.
  • The Mona Lisa challenge has 100,000 cities. On 48 cores, the tour was within 0.1 % of the proven lower bound after 174 seconds and 0.051 % after 12 minutes. After about 6 hours, it was 0.0286 % above the proven lower bound, equivalent to 0.027 % above the best known tour.

The service platform has been built and is not yet in customer use.

Part 5

Spiritia Online and its world simulation

Server prototype

A server for a massively multiplayer game with a world that runs everywhere at once. Plants, weather, animals and fire are simulated across a 16 × 16 km map every second, at a resolution of 1 m, on one AMD EPYC 9455. The same engine can simulate large living environments without the players.

  • 150,000 simulated players and 4.2 million animals. In a run of 60 seconds, the slowest simulation cycle took 47.4 ms against a budget of 50 ms. This measurement predates the network stream.
  • 50,000 simulated players and 4.2 million animals with the network stream enabled. The slowest measured simulation cycle was 37.4 ms on 40 cores.
  • About 53.4 kbit/s per player on average with a view radius of 80 m. The total for 50,000 players was 2.67 Gbit/s.

The server works with simulated players. A game client has not been built.

Artificial intelligence

CPU inference and research into learning

These are our measurements on server CPUs, without GPUs. Each project is described at the stage it has reached.

Part 6

CPU inference engine

Working software

Our engine runs large public language models on CPUs and produces reproducible results in the tests below.

  • About 37 tokens per second with Ternary Bonsai 2 27B, a public model with 27 billion parameters, using 42 CPU cores and no GPU. Its vendor reports roughly 28 to 47 tokens per second on laptop GPUs with an earlier build.
  • 138, 256 and about 560 tokens per second with the Ternary Bonsai models of 8 billion, 4 billion and 1.7 billion parameters, respectively, on 42 CPU cores without a GPU. The vendor reports 76, 120 and 229 on an Apple M4 Pro laptop GPU.
  • Reproducible output. The same input produced the same output, bit for bit, under the processing conditions we tested.

Vendor rates come from different hardware and test conditions. Their model cards give the reference figures for 27B, 8B, 4B and 1.7B. This is not a controlled comparison.

Our own inference engine, with no external model runtime. The public models retain their own licences.

Part 7

Language model research

Research in progress

Models that acquire a vocabulary from raw text and train on CPUs. The research includes learning a new fact while preserving earlier responses in the tests below.

  • 2.5 seconds to acquire a vocabulary from 1.9 GB of raw text.
  • 2 h 54 min for a complete training pass over 2.1 million stories on one CPU, without a GPU.
  • A new fact from one story. Answers changed for 2 of 3 related prompts. None of the 48 unrelated answers changed.

Produces short stories. In blind tests, logic and consistency fall short of the comparison models of the same size.

Part 8

Learning with evidence

Research in progress

A learner built to associate answers with their sources, preserve earlier learning and decline an answer when evidence is missing.

  • All 4,382 earlier items checked remained unchanged after new learning.

A research prototype tested at small scale.

Neuroscience

Brain research down to individual cells

We modelled the areas and cell types listed here, usually at the level of individual cells. We ran the models and compared the results with published measurements. Eight papers give public examples of this work.

Part 9

Brain models and research

Research in progress

The models, source code and results for the areas and cell types below. This includes the working models behind the eight neuroscience papers and our unpublished work on intelligence, reasoning and language.

Research material with eight public preprints. The working files for papers 07 to 14 and the unpublished brain models belong to this part.

Cortex

  • Cortical columns
  • PV and SST interneurons
  • VIP cells
  • Neurogliaform cells
  • Chandelier cells
  • Double bouquet cells
  • Rosehip cells
  • CCK basket cells
  • Layer 6b cells
  • Prefrontal cortex
  • Frontopolar cortex
  • Orbitofrontal value cells
  • Anterior cingulate cortex
  • Insula and von Economo neurons
  • Parietal cortex
  • Default mode network
  • Claustrum
  • Supplementary motor area

Memory and navigation

  • Dentate gyrus
  • CA3
  • CA2
  • CA1
  • Concept cells
  • Subiculum
  • Entorhinal grid cells
  • Time cells
  • Cells with persistent firing
  • Perirhinal familiarity
  • Head direction system
  • Medial septum

Thalamus

  • Relay nuclei
  • Thalamic reticular nucleus
  • Lateral geniculate nucleus
  • Matrix and anterior nuclei
  • Pulvinar

Basal ganglia and reward

  • Striatal D1 and D2 neurons
  • Fast spiking interneurons
  • Cholinergic interneurons
  • Pallidum
  • Subthalamic nucleus
  • Nucleus accumbens
  • Dopamine neurons
  • Lateral habenula

Brain state and emotion

  • Locus coeruleus
  • Raphe nuclei
  • Sleep and waking
  • Basal forebrain
  • Astrocytes
  • Circadian clock
  • Hypothalamus and hunger states
  • Amygdala: fear and extinction
  • Bed nucleus of the stria terminalis

Cerebellum

  • Purkinje cells
  • Granule and Golgi cells
  • Unipolar brush cells
  • Lugaro cells
  • Interneurons of the molecular layer
  • Inferior olive
  • Pontine nuclei

Senses and movement

  • Retina to thalamus relay
  • Primary visual cortex
  • Area MT
  • Superior colliculus
  • Saccade generator
  • Auditory pathway
  • Olfactory bulb and piriform cortex
  • Vestibular nuclei
  • Motor cortex and Betz cells

Language and thought

  • Wernicke's area
  • Broca's area
  • Semantic memory
  • Number sense

The neuroscience papers describe the published cell mechanisms. Our models of intelligence, reasoning and language remain private and are included in Part 9.

Mathematics and physics

Published proofs and work beyond them

Part 10

Model universes and the behaviour of physics

Research in progress

Simulated universes developed from a minimal starting point. The tests use criteria fixed before each run, without supplying the expected outcomes. The following results concern these models.

  • A time direction emerges. One direction has the opposite sign to the spatial directions, as in relativistic spacetime.
  • Stable spacetimes in four dimension counts. These include the 3 + 1 dimensions of our spacetime and the 9 + 1 dimensions used in superstring theory.
  • Light travels as waves. At long wavelengths its speed is the same in every direction. Nothing outruns it in the tested cases, and its two polarizations have the same speed at every wavelength. The identical propagation of the two polarizations is consistent with polarization constraints from gamma ray bursts.
  • Light and matter coexist in 61,549 universes from one family. No faster than light behaviour was found in those tests.
  • Relativistic clock behaviour in 17 universes. At three fifths of light speed, a moving clock runs at exactly 4/5 of the stationary rate. This is the time dilation factor of 5/4 predicted by Einstein's formula.
  • Antimatter in 381 universes in a separate survey of 527,958 model universes. Charged particles have antiparticles. In 5 of those 381 universes, a particle and its antiparticle annihilate into light.
  • Exactly conserved charges in 3,669 of the 61,549 universes.
  • A bounce before expansion in 326 of the 381 universes found to have a big bang. This is a different group from the 381 antimatter cases. Following the big bang into its past leads to a collapsing universe.
  • 6.9 seconds to simulate all 61,549 universes with three starts each.

Research in progress. Conservation of energy and momentum, forces at a distance and quantum behaviour remain open.

Part 11

Yang–Mills mass gap research

Research in progress

The work completed so far on the Yang–Mills existence and mass gap problem, one of the seven Millennium Prize Problems.

Unpublished working files with results proved in limited settings, including strong coupling and finite volume, where results of this kind are already known. The full problem remains open. No proof of the full problem is claimed.

Part 12

Mathematical results

Public preprints

Six freely available preprints containing full proofs, together with the research files and unpublished results behind the work. The working files for papers 01 to 06, including the phosphorylation and quantum Rabi papers, belong to this part.

  • Four families of deformed cluster maps proved integrable for every size. The two D families use their reduced form. One result settles Conjecture 7.1 in W. Kim's Integrable deformations of cluster maps of type D₂N, arXiv:2506.06182.
  • The Wang–Sontag bound is attained for every n. To our knowledge as of October 2026, this was previously known only up to n = 4.
  • D optimal designs of order 198. To our knowledge as of October 2026, this is the smallest open order. Exhaustive computation rules out 11 symmetry types. For Legendre pairs, 91 symmetry types are ruled out across the 10 lengths below 200 that remain open to our knowledge. Both sets of exclusions are unpublished. They do not resolve the open cases.

The six papers are public preprints and have not yet been peer reviewed.

Part 13

Chemistry and biological systems

Research in progress

Research on stable chemical reaction networks, flow systems that tolerate failures, transport through fungal networks and genome analysis.

The phosphorylation paper is public, and its working files are assigned to Part 12. The other research listed here belongs to Part 13 and is unpublished.

Read the work

Fourteen papers you can read now

The papers are freely available as preprints on Zenodo, which is operated by CERN. They have permanent DOIs and have not yet been peer reviewed. The links point to the latest versions. You can also find them in our Zenodo collection.

Mathematics, physics and chemistry

  1. 01The Wang–Sontag bound is attained: 2n − 1 positive steady states in the n-site sequential distributive phosphorylation system, for every nChemical reaction networks · September 2026 · doi:10.5281/zenodo.22770846
  2. 02A transfer-matrix formula and a closed recursion for the hidden symmetry operator of the asymmetric quantum Rabi model, and the sl₂ structure of its constraint polynomialsMathematical physics · September 2026 · doi:10.5281/zenodo.22770848
  3. 03Liouville integrability of the deformed D2N cluster maps for every NDiscrete integrable systems · September 2026 · doi:10.5281/zenodo.22770852
  4. 04Liouville integrability of the two-parameter deformed A2N cluster maps for every NDiscrete integrable systems · September 2026 · doi:10.5281/zenodo.22770854
  5. 05Liouville integrability of the two-parameter deformed A2N+1 cluster maps for every NDiscrete integrable systems · September 2026 · doi:10.5281/zenodo.22819874
  6. 06Liouville integrability of the two-parameter deformed D2N+1 cluster maps for every NDiscrete integrable systems · September 2026 · doi:10.5281/zenodo.22820598

Neuroscience

  1. 07Repressor regulation and release timing distinguish three coupling routes in a stochastic suprachiasmatic clock modelCircadian clock · October 2026 · doi:10.5281/zenodo.23511162
  2. 08Recruited convergence sets the sign of plasticity under a single calcium rule: converging subicular and infralimbic inputs to the bed nucleus of the stria terminalisSynaptic plasticity · October 2026 · doi:10.5281/zenodo.23511170
  3. 09Cortical cholinergic suppression and the conditions for neurogliaform and arkypallidal vetoes in a spiking model of the striatal microcircuitBasal ganglia · October 2026 · doi:10.5281/zenodo.23511188
  4. 10A bistable presynaptic switch and a nutrient-gated cue rule reproduce the daily activity pattern of arcuate hunger cellsHypothalamus · October 2026 · doi:10.5281/zenodo.23511201
  5. 11Velocity storage in a model of the paired vestibular nuclei requires slow commissural inhibition and operates near a symmetry-breaking instabilityVestibular system · October 2026 · doi:10.5281/zenodo.23511211
  6. 12A theory of latched physiological drives: hunger, sleep and defence with hysteresis and a terminating resetMotivational states · October 2026 · doi:10.5281/zenodo.23511220
  7. 13Why delaying eating need not reset hunger: a bounded-control theorem for a bistable hunger stateMotivational states · October 2026 · doi:10.5281/zenodo.23511290
  8. 14A delay-and-replacement test for feedback-maintained olfactory beta rhythmsOlfactory system · October 2026 · doi:10.5281/zenodo.23515097

The handover

What comes with the purchase

You receive the work at the stage described on this page, with the materials needed to examine it and continue it after completion of the deal.

Buying a part

  • Its source code, build scripts and tests
  • Its documentation, research notes, raw results, measurement scripts and verification records
  • Assignment of the relevant pending patent application, where there is one, at completion of the sale
  • Our transferable rights in its unpublished technology and materials, together with exclusive rights to use the code and documents, as specified in the contract
  • Our contractual commitment to stop work on the sold part, keep its unpublished materials confidential and retain no right to exploit it or usable copy, subject to the stated exceptions
  • A licence to any shared foundation needed to use that part, limited to use within it

Buying the company

  • Ownership of Steinborn IT GmbH through the agreed transfer of its shares, with the company retaining its assets, rights and obligations
  • All 13 parts and the three patent applications for SFA, the Software Evolver and the LLM Analyzer
  • The research files behind all fourteen papers
  • The company's websites

Investing in the company

  • An individually negotiated ownership share in the company. The amount, rights and form of the investment are agreed separately.

Exclusive ownership covers the rights we can grant. Existing licences, public papers and materials belonging to others keep their existing terms. Any such rights affecting the purchase are identified before you sign. Each private asset is assigned once. Shared dependencies are covered by the agreed licences. Handover support is included only if separately agreed.

The person behind it

You can also commission the next idea

I directed all the work on this page. If you want that ability applied to your own problem, you can engage me separately. After a sale of the company, I can contract directly. After a sale of individual parts, the work can be commissioned through Steinborn IT. Each arrangement is negotiated on its own, outside the bidding and subject to rights and exclusivities already granted.

Commissioned research

Give me the problem

You bring the question or the idea. I investigate it and build the result. The contract gives you the transferable rights and exclusive rights of use in the work produced for you.

Right of first refusal

See the next work first

For the scope and period we agree, I offer you new work before offering it to anyone else.

Exclusive retainer

Reserve my research time

For an agreed term, I invent exclusively for you. The contract gives you the rights to what I create within that arrangement. My next work in that scope goes to you.

Working together

Your team is welcome. A small warning.

I am autistic. Working with me can be a little like having Walter Bishop from Fringe in the team. I move between fields, notice patterns before I can explain them and ask questions that make sense a little later. When I go quiet, I am working. When I start talking quickly, I have probably found something. Give me a clear problem and room to think. The cow is not included. Probably.

I am offering independent research engagements, with agreed results and the freedom to organize my work. Write to sales@steinborn.it or reach me on LinkedIn.

Bidding and terms

Make an offer for what you want

  1. Choose the scope. Name one part, several parts or the whole company and state the amount in euros. For a package, say whether it must be accepted together or give separate prices for the parts. You may submit alternatives. State the proposed payment terms, any applicable VAT and how long the proposal remains open. Investment enquiries should state the proposed amount, ownership share and whether the money would buy existing shares or fund the company.
  2. Rights pass at completion. Steinborn IT GmbH and its founder transfer the rights they own in the sold work and grant the agreed exclusive rights of use. This includes the relevant patent application and the unpublished methods, code, documents, data and results covered by the contract. Payment, completion conditions and any required approvals must be satisfied before the confidential materials and rights change hands. Applications for further patents are the buyer's decision, subject to patentability and existing rights.
  3. We stop work on what we sell. We retain no right to exploit a sold part and no usable copy of its confidential materials. The exceptions are disclosed licences and any records the law requires us to retain, kept solely for that purpose. We also bind ourselves by contract to keep the unpublished work confidential and neither publish nor discuss it, except with your written permission or where disclosure is legally required. Our silence preserves the confidentiality of what you buy. Patentability still depends on the applicable law and any earlier disclosures. Published material and rights already granted to others cannot be withdrawn by this sale. If only parts are sold, unsold work stays with Steinborn IT and may continue.
  4. Shared foundations remain usable. Several parts depend on SFA. A buyer receives a licence to the foundations needed within the purchased part. A buyer of a foundation takes it subject to those licences and any existing licences identified in the contract. The foundation licence forms part of the purchase. Its duration, permitted commercial use and any rights to transfer or sublicense it are specified in the contract. Dependencies and overlapping rights are settled before signing.
  5. Offers close on Tuesday, 1 December 2026 at 23:59 CET. We select the compatible combination of qualifying offers with the highest total, subject to agreement on the contracts. A company offer competes with the total of the best compatible part offers actually received. Each part can be sold only once. For ranking an investment proposal, we use the proposed amount divided by the proposed ownership share. For new capital, that is the company valuation after the investment. It is distinct from the amount paid. If an investment proposal is selected, the company and its parts remain together, and the founder's continuing role is agreed in the contract. Qualification requires a lawful transaction, credible ability to pay and acceptance of the stated sale conditions. We may decline offers if no acceptable contract can be agreed. After the deadline, we contact the selected bidders to agree signing and completion. If a proposed deal falls through, we may consider the remaining qualifying offers.
  6. The methods stay closed until completion. There is no inspection of unpublished code, concepts or formulas before the sale. No NDA changes that. A separate NDA may cover an agreed demonstration of outputs. Demonstrations reserve nothing. Equal offers on otherwise equivalent terms favour the buyer who purchases blind.
  7. The company warrants its measurements. Steinborn IT GmbH warrants that its own figures were measured as stated. The scope, remedies and time limit are set in the contract. Vendor figures are attributed separately. It does not promise future research results, income, patent grants or performance on other workloads. The contract identifies the software builds and measurement conditions covered. The work is offered at its stated stage as of 24 October 2026.
  8. Investment terms are negotiated individually. This page invites enquiries. It provides no fixed subscription terms, accepts no investment payments and promises no return. Any investment must meet the applicable legal requirements before it is offered or completed.
  9. A bid does not complete a sale. This is an invitation to make proposals. No award, transfer or obligation to sell arises automatically from a bid or the deadline. A transaction requires an agreed contract, notarisation where required and any necessary approvals. Nothing on this page excludes rights or duties that the law does not allow us to exclude.

Send your proposal to
sales@steinborn.it

Include your name, the buyer, the parts or company share you want, the amount and any conditions.

Once a part is sold, its buyer decides where it goes next. That buyer could be your competitor.