In early access · a small number of partners
The proof

The evidence,
signed.

Name: HexStellar Multi-Model Field Test
Creator: HexStellar

Every number on this page comes from one controlled field measurement, across five independent labs' flagship models on one NVIDIA H100. Not a curated slide. The complete signed files are shared directly with companies evaluating HexStellar, under NDA.

See the results ↓ Get in touch

Signed, not public

We don't ask you to take our word for it — we ask you to sign an NDA.

The measurement below is real, signed, and independently timestamped. We share the complete files directly with companies evaluating HexStellar.

Request access →

17.2–27.4%Less energy per card

Up to −15°CCooler, per model

+167%More capacity, same latency

100%Identical output

Signed measurement

Five models. One GPU. A controlled measurement.

One NVIDIA H100 80GB, running vLLM 0.26.0. The layer on versus off, at identical load, across five independent labs' flagship models.

Capacity +167%

More capacity, one card, same latency 109.8 → 293.7 requests/sec at the same latency target — nearly three times the traffic.

Energy 17.2–27.4% Less energy per card, across five models Lowest: Google Gemma 3 27B, −17.2%. Highest: Meta Llama 3.3 70B, −27.4%.

Temperature Up to −15°C Cooler, per model 62.4°C → 47.4°C on Meta's Llama 3.3 70B — the largest drop measured.

Signed RFC 3161 Shared under agreement The complete signed files are available to companies evaluating HexStellar.

Lab · Model	Energy / card	Peak temp, off → on
Alibaba Qwen2.5-72B-Instruct	−26.9%	60.9 → 47.5 °C
Meta Llama 3.3 70B	−27.4%	62.4 → 47.4 °C
DeepSeek R1-Distill 70B	−24.8%	61.9 → 49.3 °C
Google Gemma 3 27B	−17.2%	52.6 → 43.1 °C
Microsoft Phi-4 (14B)	−25.5%	51.0 → 40.2 °C

A SEPARATE, DEEPER CONFIGURATION — PHI-4 ONLY

This is a different test setup from the five-model table above, run on Microsoft's Phi-4 alone under a deeper configuration. Over about 19 minutes on one server, starting from a draw of 482.7 W, the energy reduction rose from 38.0% to 50.5% and leveled off there. Board power: 306 W → 152 W. Peak temperature: 48.2°C → 39.2°C. This is the single largest result in the whole dataset — from a different configuration, not from re-running the standard comparison harder.

Correctness

100% identical output

Same answers, identical results, zero errors — the layer on versus off, across every model tested. The premise is narrow and practical: make the chips a company already owns do more work with less power, without touching the outputs.

Model independence

Not tied to one model, one lab, or one architecture

Five unrelated model families, five independent labs, sizes from 14B to 72B parameters — all through the same integration surface, with no per-model work and no changes to the model itself.

"Publish the shadow, keep the statue: the proof travels, the machinery stays home."

— Brayon Pieske

Request the signed files →

The measurement files are signed and independently timestamped. We share them directly with companies evaluating HexStellar; they are not published for public download.

What comes next

Inference is one front.
It is not the only one.

The measurement above is a language-model workload, because that is the one we finished first. The runtime is not built around language models — it operates wherever the same shape of problem shows up. These are the fronts it already runs on. Each gets the same treatment as the page above: one result at a time, measured, scoped, signed.

The one that shouldn't be possible

Problems the industry says need dedicated hardware.
We run them on a laptop.

NP-hard combinatorial optimisation — maximum cut, graph partitioning, set cover, exact cover, number partitioning — is the class an entire category of specialist machines exists to attack: purpose-built accelerators, some of them cryogenically cooled, priced accordingly. We reach the proven optimum on public benchmark instances of this class on a consumer laptop. No GPU. No accelerator. No cryogenics. No cloud.

The instances are public and checking a result is cheap, so this is verifiable by anyone who wants to — which is the point. Full figures go up with the write-up.

Routing & logistics

Travelling salesman and vehicle routing

Route ordering, capacitated vehicle routing, assignment and scheduling — the shapes that sit under fleet, delivery and workforce planning.

Selection under constraint

Knapsack, feature selection, allocation

Choosing the best subset when everything competes for the same budget — the arithmetic behind portfolio, capacity and resource decisions.

Scientific computing

Reconstruction from dense sensor data

Recovering structure from very large, noisy detector output — hundreds of thousands of points resolved in a single pass.

On-device

Phone and edge silicon

The same runtime, off the network. No datacenter in the loop, and nothing leaving the device.

Not named yet

Outside every category above

Measured, not published. That is as much as we will say in public today.

Figures for each front go up over the coming days, one at a time. Companies evaluating HexStellar see them first →

How this started

Found by accident.
Followed on purpose.

HexStellar wasn't the plan. It surfaced while we were building the infrastructure behind Trust Carbon Infrastructure — our other venture, a verification platform for carbon and environmental data, legally operated by Zenith Flow Innovations.

Deep in that infrastructure, we ran into a computational bottleneck that shouldn't have existed. We went after it anyway. Months of quiet investigation later, what we'd found was bigger than the problem we started with.

That side investigation became HexStellar. The benchmark on this page is what it does today.

Hex + Stellar

HEX

Six sides. It's nature's most efficient tessellation — honeycombs, carbon rings, basalt columns, snowflakes. It's also the language machines use to talk to themselves: hexadecimal is the base of every byte, every color, every memory address. Hex is the shape of the architecture itself.

STELLAR

Not a metaphor for ambition — a reference to efficiency. A star is the most efficient energy-conversion system that exists: fusion turns mass into energy at a ratio no engineered system has matched. That's the standard "Stellar" refers to. Not speed. An efficiency bar set by physics itself.

Our Position

"Performance is a moral question — every wasted joule is a choice someone made when designing a language."

We didn't build HexStellar to be a "green" language. We built it to be the most correctly engineered language. The environmental consequence — less energy, less heat, less hardware stress — is the natural outcome of getting the foundation right.

● Patent Pending

The company behind it

Trust Carbon Infrastructure

The venture we were building when HexStellar found us — and the reason we know how to hold an engineering claim to a standard that survives scrutiny.

Digital MRV & Verification Platform · Est. 2023

Trust Carbon Infrastructure is a methodology-agnostic verification platform for carbon credit and biodiversity projects, built to work under real field conditions — offline-first data collection from a smartphone, with anti-fraud safeguards like GPS-spoofing detection and cryptographic hashing of every record. It supports major methodologies including Verra, Gold Standard, CDM and LuxCS, in 17 languages.

Over two million lines of code across the web platform and the mobile app — and almost none of it borrowed. The field app runs completely without an internet connection: no third-party APIs, no external services to call, and its own navigation system rather than a hosted map. That was not a preference. The places this platform is used do not have a signal, so anything rented from the cloud simply would not run there.

That constraint is also where HexStellar came from. Building a system that had to do everything itself, on whatever hardware was in someone's hand, is what turned a computational bottleneck into a question worth chasing.

The platform was named one of five global winners of the DPI for People and Planet Innovation Challenge, out of 540 startups from 73 countries — a $100K award, judged by the Gates Foundation, BCG and JICA. It's the same discipline — auditable, reproducible, sealed against drift — that HexStellar's benchmark numbers are held to.

Trust Carbon Infrastructure is legally operated by Zenith Flow Innovations LLC, a U.S. company based in California.

2M+ lines of code Fully offline field app Own navigation · no third-party APIs Founded 2023 Verra · Gold Standard · CDM · LuxCS 17 languages $100K DPI Challenge award Zenith Flow Innovations LLC · California, USA

The evidence,signed.