Tencent Agent Trains Two Open Protein Models, HyInv and HyPLM

Tencent Hunyuan has published HyInv, a small protein-design model, and HyPLM-Flash, a protein language model, with their weights, inference code and technical reports, according to a 2 November update to its Hyra-results repository on GitHub. Both are released under the MIT licence. Tencent says neither was built by its human researchers in the usual way: they were trained from scratch by Hyra 2.0, the company’s automated research agent.

That is the real pitch. Tencent is presenting the two models less as products than as evidence that an AI agent can carry a scientific model from design to deployment, and then hand the result to other researchers. The performance figures that come with them are Tencent’s own, and nobody outside the company has yet reported reproducing them.

In a research blog post dated 1 November, the Hyra team describes the shift as moving automated research away from “drawing cards and grinding problems” and towards “upgrading tools”. A solved problem yields one answer, it argues, while a better tool benefits every task that later uses it.

HyInv designs sequences for a given protein backbone

HyInv tackles inverse folding: given the backbone of a target protein, find amino-acid sequences that should fold into that shape. Tencent’s blog calls it a basic tool for drug design, enzyme engineering and library construction. The model is a graph neural network with 1.96 million parameters, and its weights file is just 7.8MB.

According to the HyInv README, it follows the message-passing design of ProteinMPNN, the widely used model from David Baker’s lab, with a four-layer encoder and a three-layer decoder. It looks only at backbone atoms and runs on both CPUs and GPUs. Tencent says Hyra 2.0 handled the pre-training, supervised fine-tuning, reinforcement learning and inference optimisation, and also devised a two-step “draft, then rewrite” sampling method.

The company claims that, compared with ProteinMPNN, HyInv’s sequences match target structures more closely and are generated 57 times faster. The README gives the underlying timings: on a single Nvidia H20 GPU, the fast mode produced 128 designs in a median of about 25 milliseconds for a 148-residue protein and about 54 milliseconds for a 402-residue one, against about 1 second and 2.6 seconds for the reference code. Tencent says it has verified these speeds only on the H20.

HyPLM-Flash trades training data for speed

HyPLM-Flash is a masked language model with 650 million parameters that learns patterns from protein sequences. Such models are used to predict the effect of mutations, annotate protein function and support design work. Tencent’s blog says Hyra 2.0 trained it on 726 billion tokens, 0.73 times the data used for Meta’s ESM-2 650M and a fifth to an eighth of the data used for E1-600M and ESM C 600M.

The two published evaluations differ in scope. The blog says that among 12 protein language models compared, HyPLM-Flash had the fastest inference and an average rank of third across 35 tasks. The HyPLM-Flash README describes a matched test of 28 public checkpoints on 32 tasks, in which its zero-shot score on the ProteinGym benchmark was statistically tied with ESM-2 650M, ESM C 600M and E1-600M. It claims about 130,000 residues a second on one H20, which it says is 1.9 to 2.7 times faster than other 600M-class encoders. Unlike HyInv, it needs a CUDA GPU.

Hyra 2.0 itself splits work across three tiers of agents, according to the blog: an “Overall” agent that talks to human researchers, “Mission” agents that each own a research problem, and “Worker” agents that run experiments. Jobs that need large amounts of computing power still go through human approval.

Small, open models still need outside testing

The release fits a pattern at Tencent Hunyuan, which has been open-sourcing agent tooling such as its Hy-Memory 2.0 SDK and models such as the Hy-Nomas world model, while pushing agents commercially through WorkBuddy ahead of its Hy4 release. What is different here is the claim about authorship. If an agent can repeatedly train useful scientific models end to end, the bottleneck in fields such as protein design moves from engineering effort towards checking the results.

That checking is the open question. HyInv’s headline numbers come from Tencent’s own benchmarks, on one type of Chinese-market Nvidia chip. Speed is the easier claim to test; sequence quality, measured by how well computer-predicted structures match the target, is a step removed from whether a designed protein actually folds and works in a lab. Neither model has been described as tested in wet-lab experiments in the material Tencent has published.

What works in Tencent’s favour is the format. A 7.8MB model under a permissive licence, with code and a technical report, is cheap for academic groups to run against ProteinMPNN on
their own backbones. Independent benchmarks of open protein models already exist: the field has recently seen the openly released RFD4-Proteina design model from UW and Nvidia, giving outside groups a point of comparison. The faster those comparisons arrive, the sooner Tencent’s figures can be confirmed or quietly trimmed.

Maths papers come without peer review

The same repository, and the blog, go well beyond proteins. Tencent says Hyra 2.0’s HyProver workflow, “together with mathematicians”, has verified and written up proofs of two conjectures: a dimension-five case of the Petty equilateral-points problem, and a counterexample to Yau’s circle-action conjecture. On the second, Tencent is explicit that the counterexample is a closed seven-manifold that is not simply connected, and that the simply connected version of the conjecture remains open. These are papers posted by the company, not results that have been through journal peer review, and should be read that way for now.

The blog lists further tools Tencent says it will release on GitHub, including HyQEC for quantum error-correcting codes and work optimising the open-source XiangShan RISC-V processor. Until outside researchers run the weights and the proofs for themselves, the honest summary is that Tencent has shown its working and invited the community to check it.

Sources

• Tencent-Hunyuan/Hyra-results repository, GitHub (News entry, 2 November 2026)
• HyInv model README and technical report, GitHub
• HyPLM-Flash model README and technical report, GitHub
• Tencent Hunyuan, “Hyra 2.0: the next step for automated research” (research blog), 1 November 2026

Related stories

• Tencent Hunyuan Open-Sources Hy-Memory 2.0 Agent Memory SDK
• Tencent Hunyuan Open-Sources Hy-Nomas, an Interactive World Model
• Tencent Says WorkBuddy Passed 9bn Tasks as It Previews Hy4 Release