What You Still Need to Reproduce Tencent Hyra HyInv and HyPLM-Flash

At 16:50:32 Asia/Shanghai on 2026-11-02 (08:50:32 UTC), commit df852cb881e4ade1eb00c493c7db4aac1c22e744 landed on Tencent-Hunyuan/Hyra-results with the message Add HyInv and HyPLM-Flash model releases under AI4Science. Seconds later, Release hyplm-flash-weights published (16:50:59 Asia/Shanghai / 08:50:59 UTC), marked prerelease: false and draft: false. The drop is model definitions, inference code, English READMEs, technical-report PDFs, HyInv weights in-tree, and HyPLM-Flash FP32 weights as a formal GitHub Release—not a preview tag.

Open inference and open weights answer a deployment question. They do not answer a full training-reproduce question. The rest of this guide maps what the commit, the Release page, and the two English READMEs actually pin: environments, precision paths, seed and device rules, and the training gap that remains. We have not run these models end-to-end or remeasured the throughput and accuracy figures printed in the READMEs and technical reports; those numbers stay labeled as self-reported when cited.

Related on-site context: Hunyuan Opens Hy-Nomas: Interactive Video World Model on a 4090, Tencent Opens Hy-SimulMT — Download Isn't Instant Interpreting, Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face, and MiniMax M3.1 Flash Open Weights Are on Hugging Face. Those pieces sit on other Hunyuan or open-weights lines; this article stays on Hyra AI4Science inverse folding and protein language-model inference.

What the commit put under AI4Science

Per the commit message, both folders ship model weights (HyPLM-Flash’s large weights go through the Release), model definition and inference code, a technical report PDF, and an English README. Author and committer timestamps on the commit object are 2026-11-02T08:50:32Z and 08:50:44Z—16:50:32 and 16:50:44 Asia/Shanghai. The paths are AI4Science/hyinv/ and AI4Science/hyplm_flash/.

HyInv and HyPLM-Flash are credited in the READMEs as trained by the research agent Hyra 2.0. License files in both trees are MIT. That covers the released packages as documented; it does not invent a training-data dump that the repository does not ship.

Release hyplm-flash-weights: formal FP32 shards

Release API fields match the HTML tag page:

• tag_name: hyplm-flash-weights
• name: HyPLM-Flash weights
• prerelease: false · draft: false
• published_at: 2026-11-02T08:50:59Z ＝ 2026-11-02 16:50:59 Asia/Shanghai

Assets:

Release body: FP32 weights, two shards; download both, cat into model.safetensors, then verify sha256 ce7dcd4fc473dc334d869718a9a31149573893795db03c45efeb396e8b180ab1. The same flow lives in AI4Science/hyplm_flash/weights/download.sh. Per-shard digests on the assets API are separate from that merged checksum; verify the merged file against the Release body value.

HyInv: 1.96M inverse folding on CPU or GPU

HyInv’s README describes a 1.96-million-parameter graph network for protein inverse folding (backbone → sequence), ProteinMPNN-style message passing (4-layer encoder, 3-layer decoder), seeing only N / Cα / C / O backbone atoms, runnable on CPU and GPU. Released weights are the structure-reward reinforcement-learning version; the documented sampler runs 16 rounds (encode once, 8 blind drafting rounds at temperature 0.6, then 8 rewrite rounds at temperature 0.34).

Weights in-tree. Layout points at weights/model.safetensors as FP32, 7.8 MB in the README. The GitHub Contents API size for that file is 7843508 bytes (≈ 7.48 MiB). Both figures describe the same artifact two ways—there is no third volume.

Install surface. Python 3.10+, PyTorch, safetensors, numpy; accelerated CUDA path adds triton via pip install -e '.[fast]'. CLI shape from the README: python -m hyinv design --pdb structure.pdb --n 128 --seed 0 --out designs.fasta, with options for chain, fixed residues, temperatures, --device cpu|cuda, --fast off|exact|bf16, and --weights.

Reproducibility pins (README Notes)—these are the hard constraints. Outputs match only when seed, batch size (default 64, auto-tightened by length), device type, and PyTorch version stay the same. Each batch has its own stream: batch k uses seed+k, so changing batch size changes results. CPU and GPU random streams differ, so sequences from CPU versus GPU are not identical even with the same seed. That CPU≠GPU rule is easy to miss when a laptop smoke test is compared to a GPU lab run.

--fast modes. Default is off: FP32 reference on CPU or GPU. exact: FP32-level acceleration (TF32×3 accumulation, erf GELU); with matching arguments, sequences are almost always identical to off, with rare flips when probabilities nearly tie. bf16: BF16 operands, FP32 accumulation, tanh GELU—fastest path, numerically different; same seed does not guarantee the same sequence as off, though the path is reproducible with itself. Triton kernels were verified on a Hopper-class H20 (sm_90); other architectures “should compile” per the README and have not been verified there. First-call compile and CUDA-graph capture costs are documented separately from steady-state latency.

README and technical-report millisecond timings (for example bf16 ≈ 25 / 54 ms for 148 / 402 residues on one H20 under stated conditions) are self-reported. We have not remeasured them.

HyPLM-Flash: 650M masked PLM, CUDA required

HyPLM-Flash’s README describes a 650-million-parameter masked protein language model. Installation needs a CUDA GPU; Hopper GPUs such as H20 or H100 are recommended. The FP32 weights are the Release shards above (~2.61 GB), fetched with bash weights/download.sh (curl; GITHUB_TOKEN only if rate-limited). pip install -e . pulls torch, safetensors, triton, and flash-attn; FlashAttention-3 is used when present, otherwise FlashAttention-2.

Precision table (README). Default precision="int8" targets representation extraction, zero-shot scoring, and structure / function prediction; README self-reports about 130,000 residues/s on an H20. precision="bf16" self-reports about 90,000 residues/s and is recommended for training supervised regressors on near-identical variants (for example AAV and GB1 in FLIP), because INT8 representations of same-protein variants differ by only about 1% in the README’s wording. Under INT8, zero-shot / structure / function / likelihood metrics are said to sit within 0.01 of BF16. Again: document claims, not our measurements.

Inference surface: HyPLMEncoder.from_pretrained("weights") (INT8 by default), encode / logits, optional multi-GPU helpers, and CUDA-graph paths for low-latency single-protein work. Single-sequence length up to 8,190 residues is documented; packing uses batch_tokens (default 32768).

ProteinGym zero-shot Spearman 0.409 and the “best long-range contact precision” line in the README are self-reported evaluation copy. We have not reproduced that eval grid.

Open inference ≠ full training reproduce

What this repository open-sourced for these two models, on the evidence of the commit and READMEs:

• Model definition and inference code
• Technical-report PDFs
• HyInv FP32 weights in-tree (~7.8 MB / API 7843508 B)
• HyPLM-Flash FP32 weights via formal Release (~2.61 GB, two shards)
• MIT license text

What is not established as a turnkey training reproduce from those same pages:

• Full training corpora and exact data recipes beyond the README’s high-level comparisons (for example HyPLM-Flash’s “0.73× the data of ESM-2 650M” phrasing)
• End-to-end training scripts, schedules, and reward pipelines as a complete public replay
• A guarantee that every README latency or ProteinGym figure will match a local stack without matching hardware, software, and seeds

So: deployable inference packages with published weights, plus documented reproducibility rules for sampling and encoding—not “clone the repo and the entire training run falls out.”

Operator checklist

1. Clock the drop. Commit and Release cluster around 16:50 Asia/Shanghai on 2026-11-02 (08:50 UTC); Release is formal (prerelease: false).
2. Pick the right tree. Inverse folding → AI4Science/hyinv. Masked PLM → AI4Science/hyplm_flash + Release download.
3. HyInv reproducibility. Lock seed, batch size, device type, PyTorch version; remember seed+k per batch and CPU≠GPU streams.
4. HyInv --fast. off is the FP32 reference; exact ≈ reference sequences; bf16 is another sample from the same distribution, not a bit-match to off. Hopper H20 verified for kernels; other GPUs unverified in the README.
5. HyPLM-Flash hardware. CUDA required; Hopper recommended. No README claim that CPU inference is supported.
6. HyPLM-Flash precision. Default INT8 for representations / zero-shot / structure-function style work; BF16 when fine-tuning regressors on near-identical variants.
7. Verify weights. HyPLM-Flash: merge shards and check sha256 ce7dcd4fc473dc334d869718a9a31149573893795db03c45efeb396e8b180ab1. HyInv: README 7.8 MB vs API 7843508 B for the same model.safetensors.
8. Promo metrics stay self-reported. Throughput, millisecond tables, ProteinGym 0.409, and similar lines read as README / technical-report claims; we have not remeasured them.
9. License. Both packages document MIT; redistributions retain the LICENSE files.
10. Topic. The subject is Hyra AI4Science protein tooling—not Hy-Nomas, not Hy-SimulMT, and not a claim that training is fully public.

Bottom line

Tencent’s Tencent-Hunyuan/Hyra-results tree, at 16:50 Asia/Shanghai on 2026-11-02, opened HyInv (1.96M, CPU/GPU, tight seed/batch/device/PyTorch rules) and HyPLM-Flash (650M, CUDA/Hopper-oriented, INT8 default vs BF16 for near-variant regression) with inference code, reports, and weights—including a formal HyPLM-Flash Release of about 2.61 GB. That is enough to stand up documented inference paths. It is not automatic proof of a complete training reproduce, and we have not measured local speed or accuracy against the README figures. Live run reports can go through the site-owner contact in the footer.