Open in tt-awesome →

tt-bio

affiliated★ featured
by moritztng · Python · MIT · 126⭐ · Jan 31, 2026

Boltz-2 biomolecular model for drug discovery on Tenstorrent Blackhole. Supports single-card and multi-card configurations — QuietBox (4×) and Galaxy (32×). Approaches physics-based FEP accuracy at 1000× the speed.

LATEST v0.27.0 2026-11-06T03:21:39Z Release notes ↗
4 previous releases
v0.26.1 2026-10-29T10:04:00Z
v0.26.0 2026-10-29T01:14:33Z
v0.25.0 2026-10-28T02:21:24Z
v0.24.0 2026-10-27T18:05:56Z
See all releases on GitHub ↗
# Changelog

All notable changes to TT-Bio are recorded here. Versioning is [SemVer](https://semver.org);
releases are cut from a commit that has passed the on-hardware test suite (see `RELEASING.md`).

## [Unreleased]

### Added
- **BindCraft 2 on the benchmarks page.** 352 s per 140-iteration design trajectory against PD-L1 on one Blackhole
  p150a, against 44.4 s on a B200, 52.0 s on an H200 and 119 s on an A100, each NVIDIA card on BindCraft 2's fastest
  shipped option. A Galaxy Blackhole delivers 1.8x a DGX B200's designs per dollar of purchase price.
- **A PPI grade says when it could not score the binder's pose.** The two binder-pose RMSDs under
  `ppi_refolding` align on the target's absolute conditioned coordinates, so a design whose target was pinned with
  `C_CRDGRP` (an internal coordinate group, RFD4's recommended binder recipe) has nothing to align on and the `ppi`
  criterion comes back `pass = 0` with `all_metrics_present = 0`. That is upstream's behaviour and it reads exactly
  like a design that was scored and failed. `tt-bio evaluate` now names those designs and the columns it could not
  compute, and `docs/rfd4-design.md` says which conditioning to use if you intend to grade a binder's pose.
### Fixed
- **An RFD4 grade scores the designed chain again.** Every metric defined on the chain of interest
  (`refolded_target_aligned_binder_rmsd`, `refolded_target_backbone_aligned_binder_rmsd`) was computed over an
  empty atom subset, because the evaluate path matched the oracle's chain label against a chain instance id it
  synthesised itself: `B_1` against the `B` the oracle wrote. The subset is no longer empty, and a confidence JSON
  written with either spelling now resolves to the same chain. A refold also inherits its design's `C_CRD` and
  `SA_CRDSRC` masks, which an oracle's plain output structure does not carry, so metrics aligned on them are
  scored rather than dropped.

## [0.27.0] - 2026-11-06

### Added
- **Mol-1 on the benchmarks page.** 49.6 s per 512-residue fold (20 recycles, 200 sampling steps) on one Blackhole
  p150a, against 18.0 s on a B200, 24.0 s on an H200 and 50.1 s on an A100, each NVIDIA card on Mol-1's fastest shipped
  option. A Galaxy Blackhole delivers 5.1x a DGX B200's folds per dollar of purchase price.
- **RFD4-Proteina on the benchmarks page.** 3.74 s per design on one Blackhole p150a at batch 40, against 1.12 s on
  a B200, 1.48 s on an H200 and 3.50 s on an A100, each NVIDIA card at its own best batch. A Galaxy Blackhole
  delivers 4.2x a DGX B200's designs per dollar of purchase price.
- **Mol-1 (`--model mol1`).** Molence Bio's Protenix-derived structure model folds protein, DNA, RNA, ligand and ion
  complexes on the Protenix engine. It reads upstream's job-list JSON unchanged, writes upstream's per-seed output
  layout with its summary and full-data confidence files, and defaults to upstream's 20 recycles, 200 steps and 5
  samples. For Protenix and Mol-1 job JSON, `--use_seeds_in_json` run
drug-discovery blackhole inference biology multi-card
blackhole quietbox galaxy