Open source

Biomolecular AI at
a fifth of the cost

TT-Bio runs the leading structure prediction, design and affinity models on Tenstorrent hardware. On OpenFold3 and Boltz-2, one Blackhole Galaxy has higher throughput than an NVIDIA DGX B200, for under a third of the price.

5×
throughput per dollar
of purchase price, against a DGX B200
One
codebase
optimized from silicon to serving
1:1
accuracy parity
measured against every reference

Benchmarks

Throughput, every model, every system

Predictions per hour on the same 512-residue protein, each stack on its own shipped settings, then divided by what the hardware costs.

The GPUs run Anthropic's open kit for these models in exact mode where it has the model, and the authors' own release where it does not. Methods

Galaxy BlackholeNVIDIA, Anthropic kitNVIDIA, upstream

Every number, and how it was measured →

Why it costs less

These models want many small processors, not a few huge GPUs

DGX B2008 large GPUsGalaxy32 smaller processors

Small models, many small processors

A biomolecular model is 2 to 3 GB of weights, about 1 % of a B200's memory, so one prediction leaves most of a big GPU idle. A Galaxy gives every prediction its own processor: 32 run at once.

HBMHBMHBMHBMHBMHBMHBMHBMComputeHBM buys memory speedExpensive, and sold outBio models are limitedby compute

Compute-bound, so no HBM

The pairformer does N³ work on N² data: about 676 calculations per byte at 512 residues, past the point where any GPU waits on memory. Blackhole pairs GDDR6 with large on-chip SRAM instead of HBM.

Off the HBM allocation

The memory makers report their 2026 HBM output sold out. Blackhole does not use it, so you can actually buy it.

Inference and training

The same stack trains OpenFold3-family models, on the same hardware.

Coverage

Every model, one codebase

Every model runs on the same stack, from the kernels on the chip to the serving layer. Speed up one part and every model gains.

Structure prediction

OpenFold3AlphaFold3-family all-atom
Boltz-2Structure and binding affinity
OpenDDEAntibody-antigen and multi-chain complexes
Protenix-v2All-atom complex prediction
RoseTTAFold3All-atom biomolecular assemblies
ESMFold2Rapid single-sequence folding
OpenBind-0Protein-ligand co-folding
Protenix-v1Lighter all-atom complex prediction
Mol-1Protenix-family complex prediction from Molence

Training

OpenFold3Fine-tune or train on your own data

Binder design

BindCraft 2Design loop with AlphaFold2 on the card
RFdiffusion3All-atom design and scaffolds
RFD4-ProteinaBinders and scaffolds through a latent flow
BoltzGenBinders against any target
PXDesignBinder backbones from a target
AF2-IGScores the binders you designed

Binding affinity

Nesso-1Affinity scalars, no structure

Embeddings

ESM-CSequence embeddings, to 6B
SaProtStructure-aware embeddings

Accuracy

The same answer as the reference

Every release checks each model against its official reference implementation, run the way people use it: with an MSA wherever the model takes one.

The bar moves, so we measure it

A single reference run is not a target. The reference itself varies between seeds, so we measure that spread and land inside it.

Checked three ways

Submodule activations by correlation, end-to-end outputs on identical noise, and published benchmarks reproduced.

Every release, every model

A release ships when every model's checks pass. Per-target results, thresholds and reproduction commands are in the repository.

Try it on real targets, free

JapanFold runs TT-Bio on Tenstorrent Galaxy servers as a public service. Fold, design or score from the browser, the API or your AI agent.

Contact

Talk to the people who build TT-Bio

Bringing these models onto your own hardware, need one TT-Bio does not run yet, or want to train on your own data? Tell us what you are working on, and we will reply by email.

To run the models without hardware, use JapanFold. For a bug, open an issue.

Sent, thank you.

We will reply to the address you gave.

Tenstorrent, Blackhole and Galaxy are trademarks of Tenstorrent Inc. NVIDIA, DGX, H200, B200 and A100 are trademarks of NVIDIA Corporation. Hardware names are used to identify the systems measured. Line drawings: barstar bound to barnase (PDB 1BRS) and a BindCraft 2 binder on PD-L1 designed on Tenstorrent, drawn from their C-alpha traces.