One format with a fidelity dial
A single package, served at the fidelity the hardware allows.
- CompactEdge GPUs
- BalancedClusters
- Bit-exactHigh-memory servers
Lab phase · Benidorm, Spain · Founded 2026
The verifiable layer between AI models and hardware.
SpiralThink builds a model format and runtime that turns a large model into a verifiable unit you can move, shard and serve on heterogeneous hardware — from legacy servers to current datacentre GPUs — with proof that what runs is exactly what was shipped.
01 Vision — where we are heading
Thesis · not a result
A north star for how large models are stored, moved and served.
This is the original thesis behind SpiralThink. It describes where the work is heading, not what has been demonstrated. Measured results are listed separately under Milestones.
A universal representation
A universal geometric representation of model weights that allows exact reconstruction, at ratios that grow with model scale.
Large models on any pool of machines
Very large models served on whatever machines are available — edge devices, legacy servers, compute farms, cloud — distributed like files over a peer-to-peer network, but with verification.
Lineage that travels almost for free
Fine-tunes and derived models shipped as compact, verifiable recipes, so a model's lineage can travel with it at almost no extra cost.
Compute sovereignty for Europe
Making full use of the hardware Europe already has and will have, across vendors and generations, so serving large models does not depend on a single class of machine.
Efficiency per token
Efficiency measured in the memory and energy freed per token served — a quantity that can be checked, not a headline.
How to read this section: everything above is aspirational. Figures we publish elsewhere on this site are labelled MEASURED, DECIDED or PROJECTED.
This section is aspirational and does not describe current results. See the disclaimer in the legal notice.
02 Method
One format. Verification built in.
The same model package adapts to the machine it runs on, and carries what is needed to prove it is intact.
Verification built in
Integrity is part of the format, not an add-on.
- Per-tensor hash
- Merkle root per package
- Reproducible reconstruction across nodes
Weights stay compressed on the GPU
During execution the weights remain compressed in GPU memory. The same package can therefore be served on machines with very different memory budgets.
- StoredCompressed
- Moved between nodesCompressed
- Executed on the GPUCompressed
Every figure carries its status
We label every number we publish, so it is always clear what kind of claim it is.
- Measured
- Obtained on real models and hardware, and reproducible.
- Decided
- A design choice we have committed to.
- Projected
- An estimate or target that has not been measured yet.
The three are never mixed in the same figure.
03 Milestones
What has been measured so far.
A dated record of results. Each entry is labelled with its status.
-
Initial thesis and preprint v0.9
The research thesis is written up as preprint v0.9 and becomes the basis of our patent work.
-
v0.9.0 reference release on PyPI
Signed builds (sigstore) and 99% test coverage. Open decoder (CC-BY) and open encoder (Apache-2.0).
-
Bit-exact reconstruction on real models
156/156Verified block by block on real models. TinyLlama: 156 of 156 blocks.
-
Multi-platform
P40 → H100One kernel build running from NVIDIA Pascal (P40) to Hopper (H100).
-
Gemma-2-27B on H100
−0.25%Perplexity 0.25% lower (better) than native bf16, on WikiText-2.
-
Multi-model
117M – 27B4 model families, from 117M to 27B parameters, with no calibration data.
-
Memory
2.58 GB2.58 GB GPU memory peak in an isolated Gemma-27B layer benchmark, with weights kept compressed on the GPU (the full model in bf16 needs ~54 GB).
-
Production kernels in the runtime
Production kernels integrated into the runtime for Hopper and single-stream decode.
-
Provenance as a verifiable recipe
121 KBA fine-tuned GPT-2 shipped as a 121 KB verifiable recipe (0.008 bits per weight). 50 of 50 tensors verified against a Merkle root.
Figures refer to the specific models, hardware and benchmarks stated in each entry.
These are lab measurements, not product commitments or performance warranties. See the disclaimer in the legal notice.
04 Why now
Three pressures are converging.
Compression built around verifiable protocols: efficiency without losing auditability.
Cheaper inference
Serving large models is limited by memory and energy. Using the hardware already deployed matters as much as buying new hardware.
Compute sovereignty
Organisations want to run models on infrastructure they control, across vendors and hardware generations.
Traceability in production
Regulation such as the EU AI Act raises the bar on documenting which model runs where. The format is designed to support these traceability and documentation requirements.
Multi-platform. Multi-model. Multi-capacity.
Not another agent — the layer agents, models and auditors rely on.
05 Model · How we work
Open where it matters for trust. Licensed where it matters for industry.
Open reference layers
The reference decoder (CC-BY) and encoder (Apache-2.0) are open, so anyone can verify a package independently.
Industrial encoder and GPU kernels
The industrial encoder and the GPU kernels are available under licence.
Partners
We are open to research partners, public R&D programmes and teams that distribute or serve large models.
Ways to work together
Selective, scoped engagements:
- Architecture reviews
- Technical due diligence
- Pilot integrations
06 Contact
Talk to us.
For research collaboration, R&D programmes, technical reviews or pilot discussions, write to us.
Emailhello@spiralthink.com
LinkedInlinkedin.com/company/spiralthink
Company details
SpiralThink, S.L.
NIF B66136813
Avda. Montbenidorm 6
03502 Benidorm (Alicante), Spain
SpiralThink, S.L.
NIF B66136813
Avda. Montbenidorm 6
03502 Benidorm (Alicante), Spain