Flagship service · Available now

Wider residual streams.
Zero stability tax.

mHC — Manifold-Constrained Hyper-Connections — restores what unconstrained hyper-connections break: the mixing matrix is projected onto the Birkhoff polytope, so signal propagation stays bounded by construction. Measured at 27B scale: −0.021 training loss, +2.3 pp BBH, 6.7% end-to-end overhead. We make it run on your cluster.

Public mathematics. Private engineering. · arXiv:2512.24880

Kickoff within 10 business days · GBP / EUR / PLN invoicing
Overhead at n=4
0%
Training loss · 27B
0
BBH · 27B
+0 pp
Validated scale
3–27B
01 · The problem

Hyper-connections widen the residual stream — and quietly bill you for it.

Expanding the residual stream to n parallel copies with learnable mixing lifts quality. Unconstrained, it also breaks the two properties that made residual networks trainable in the first place.

Identity mapping lost

The clean gradient path disappears

In a residual network, xl+1 = xl + F(xl) guarantees an unobstructed gradient route through depth. Arbitrary mixing matrices destroy that guarantee — products across layers drift away from identity.

Signal amplification

Gradient spikes up to ~3×10³

Composed unconstrained mixing amplifies norms multiplicatively. In the published measurements, HC training exhibits gradient-norm spikes on the order of 3×10³ — the kind of instability that ends a multi-week run.

The memory wall

n× activations, naively

An n-wide stream means n× residual activations. Without kernel-level engineering, that alone prices the method out of production — which is precisely where most teams stop.

02 · The mathematics

Constrain the mixing to the Birkhoff polytope. Everything else follows.

These are published, checkable facts — not our claims. We cite them because the audience we work with checks.

Hres ∈ 𝔅ₙ

Doubly stochastic mixing

Sinkhorn–Knopp (~20 alternating normalizations) projects the residual mixing matrix onto the Birkhoff polytope: non-negative entries, every row and column summing to 1.

‖M₁M₂…ML‖₂ ≤ 1

Bounded by construction

Doubly stochastic matrices are closed under composition and non-expansive in ℓ₂ — a convex combination of permutations cannot amplify. Depth-wise products stay bounded; the identity-mapping property is restored, not approximated.

6.7% @ n=4

Infrastructure-priced

With fused kernels, selective recomputation and pipeline-parallel scheduling, the published end-to-end training overhead is 6.7% at n=4 — the difference between a paper result and a production decision.

Sources: Xie, Wei, Cao et al., arXiv:2512.24880 (mHC) · Zhu et al., arXiv:2409.19606 (Hyper-Connections) · He et al., CVPR 2016 (residual learning).

03 · The evidence

Measured on DeepSeek-V3-based configurations, 3B to 27B.

Gradient norm across training log₁₀

Illustrative visualization of the published stability behaviour — HC spikes to ~3×10³; mHC remains bounded. Not raw paper data.

Loss · 27B
−0.021

Lower final training loss versus baseline at 27B — at a scale where a third decimal is real money.

BBH · 27B
+2.3 pp

Downstream reasoning gain (BBH) reported alongside improvements on further benchmarks.

Overhead · n=4
6.7%

End-to-end training overhead with the full engineering stack applied — the number that turns research into a line item.

04 · See the constraint work

Sinkhorn–Knopp, live in your browser.

A random positive 4×4 matrix is alternately row- and column-normalized until it lands on the Birkhoff polytope. Watch the sums converge to 1 — this is the projection that keeps an n-wide residual stream bounded. In production it runs in ~20 iterations, fused, at negligible cost.

Row / column sums: farclose= 1.000

Sinkhorn–Knopp · 𝔅₄ iteration 0 · max dev
Press “Run projection”.
05 · How an engagement runs

Four stages. Each one ends in a number, not a slide.

1

Call — 30 min

Free and technical. Your stack, scale, precision policy, parallelism. We tell you plainly whether mHC is worth your time.

2

Demo — 5 days

The flagship entry point. mHC running on a proxy of your configuration, instrumented side-by-side with your baseline. Go/no-go report.

3

Pilot — 3–4 weeks

Integration in one production training pipeline, A/B at an agreed token budget, full telemetry. Code IP transfers to you.

4

Production — 8–12 wks

Multi-node rollout, monitoring, regression suite, team enablement. 90 days of engineering care.

06 · The offer — available now

You've read the paper. The question is whether your cluster can collect.

This page won't teach you what doubly stochastic mixing is — you know. It tells you exactly what you get, in what time, for what price. Fixed scope, fixed price, three currencies.

Flagship entry

mHC Demo

£5,000
€5,900 · 25 900 zł · net
5 business days
  • Live mHC run on a scaled proxy of your configuration — your dims, your precision policy
  • Gradient-norm and loss instrumentation, baseline vs mHC (n=4), side by side
  • Memory and throughput delta — measured, not estimated
  • Written go/no-go report with the expected effect at your target scale
  • 90-minute technical debrief with your team
  • 100% of the fee credited against the Pilot
Book the Demo
Most chosen
Proof on your pipeline

Pilot

£15,000
€17,500 · 77 500 zł · net
3–4 weeks
  • mHC integrated into one production training pipeline
  • Constrained mixing layer with a fused kernel path
  • A/B training runs at an agreed token budget, full telemetry
  • Stability report: spike analysis, loss delta, benchmark evals
  • Code IP transferred to you · 30 days of support
Start the Pilot
At scale

Production

from £38,000
€44,500 · 197 000 zł · net
8–12 weeks
  • Multi-node rollout, pipeline-parallel compatible
  • Mixed-precision kernel path, memory-optimised
  • Monitoring, alerting and a regression suite around the training loop
  • 2-day team enablement plus operational runbooks
  • 90 days of engineering care
Scope Production
What stays with us

You get the numbers on your workload and the code you paid for. Our kernels, integration playbook and tuning heuristics are not published anywhere — including this page. The mathematics is public; the engineering is the product. NDA as standard, in both directions.

Add-on · Team workshop

One day, on-site or remote: the mathematics, the failure modes, the operational practice of n-wide residual streams. For research and infra teams.

£1,800 · €2,100 · 9 300 zł · net

Scope Demo Pilot Production
Live run on a proxy of your config
Instrumented A/B at token budget
Production pipeline integration
Fused kernel path
Multi-node · pipeline parallel
Monitoring & regression suite
Written report & debrief
Code IP transfer
Support windowQ&A 14 d30 d90 d

Net prices. Invoicing from a UK Ltd (Companies House № 16019829) in GBP, EUR or PLN — your choice, fixed at signature. Staged payments. Kickoff within 10 business days. We run a limited number of concurrent engagements. First call is free and commits you to nothing.

07 · Straight answers

Asked by people who read papers.

Is mHC CSS Ltd's technology?

The mathematics is published research by DeepSeek-AI (arXiv:2512.24880) — we cite it because our audience checks. What we sell is the engineering that makes it hold on real clusters: kernels, integration, telemetry, operational practice. That part is ours, and it stays ours.

Why doubly stochastic, specifically?

Because the Birkhoff polytope is closed under multiplication and non-expansive in ℓ₂: a product of doubly stochastic matrices is doubly stochastic and cannot amplify a signal. That restores the identity-mapping property residual networks depend on — as a structural guarantee, not a regularizer you hope holds.

What do you need from us for the Demo?

An NDA (ours or yours), your model configuration (dims, depth, precision policy, parallelism), and a technical contact. Compute can be yours or ours — agreed on the first call. No weights, no data leave your perimeter.

Does it work with our parallelism setup?

The published results were obtained in a pipeline-parallel training stack, and the method is compatible with standard tensor/pipeline schemes. Whether it holds in your exact topology is precisely what the Demo measures before you commit to more.

What if the Demo says “no-go”?

Then you've spent a fixed, known amount to avoid a much larger mistake — and you keep the report and the numbers. If it says “go”, 100% of the Demo fee is credited against the Pilot. Either way you decide on data, not on a deck.

Contract, currencies, invoicing?

CyberSentinel Solutions Ltd, a UK company (№ 16019829). Fixed-scope, fixed-price statements of work, staged payments, invoicing in GBP, EUR or PLN. NDA as standard, both directions.

08 · Contact

Thirty minutes. Bring your config.

The first call is technical, free, and honest — if mHC isn't worth it at your scale, we'll say so and save us both the paperwork.

Write directly
hello@cyberssl.co.uk

Reply within 1 business day.

CyberSentinel Solutions Ltd
Companies House № 16019829
Bristol · London · cyberssl.co.uk
Invoicing: GBP / EUR / PLN

Opens your e-mail client addressed to hello@cyberssl.co.uk. NDA available before any technical detail is exchanged.

mHC Demo — flagship entry · £5,000 net