Materials
Generative models, machine-learned interatomic potentials and the libraries they sit on. The question that runs through them: can the generated structure be made, and does the leaderboard potential hold on a real system.
A diffusion-based generative model that proposes inorganic crystal structures across the periodic table and can be fine-tuned to condition generation on chemistry, symmetry and properties.
Weights: Base and property-conditioned checkpoints are published on Hugging Face (not gated, license: mit) and via Git LFS; the README notes they were re-trained with this repo and are not identical to the paper's checkpoints, and the 4.4k ordered + 117.7k disordered ICSD structures used in the paper cannot be shared for license reasons.
Where users get stuck
- Conditional generation does not hit the requested target: users report band-gap or space-group conditioning returning off-target results (near-zero gaps, all P1). #213#156#67
- Training or fine-tuning on one's own data and property labels, including multi-GPU setup. #128#82#61
- Getting usable samples out of the pipeline at all: generation, post-training sampling and relaxation errors. #62#93#136
The choice it embodies: Code and weights are MIT and ungated, but the paper's evaluation relied on DFT and licensed ICSD data; the repo's default evaluation instead uses the MatterSim ML force field, and the README itself warns results may differ from DFT and should be confirmed with DFT. Openness of weights does not by itself make the published novelty and stability numbers reproducible.
Benchmark and reality: The Nature paper's single synthesized candidate, TaCr2O6, came out as a compositionally disordered version of the predicted ordered structure, with a measured bulk modulus of up to 169 GPa against a 200 GPa target (the paper's 'within 20%'); a 2026 Materials Horizons paper (Juelsholt) argues by crystallographic analysis that this phase is identical to Ta1/2Cr1/2O2, first reported in 1971 and included in MatterGen's training data. This echoes Cheetham & Seshadri's 2024 critique of Google's GNoME, which found 'scant evidence' of compounds meeting novelty, credibility and utility.
The question it answers: When your headline claim is 'we discovered a new material', how do you define 'novel' so it survives a domain expert, and what validation gate would you require before announcing?
Sources: huggingface.co/…/microsoft/mattergen · raw.githubusercontent.com/…/HEAD/README.md · nature.com/…/articles/s41586-025-08628-5 · doi.org/…/10.1039/D6MH00268D · europepmc.org/…/MED/42007796 · doi.org/…/10.1021/acs.chemmater.4c00643 · Company profile →
A library for training and running MACE higher-order equivariant message-passing interatomic potentials, plus pretrained foundation models (MACE-MP, MACE-OFF, MACE-MH and others).
Weights: Weights are downloadable without a form, but licenses differ by model: MACE-MP-0a/0b3 and MACE-MPA-0 are MIT, while MACE-OMAT-0, MACE-OFF23, MACE-MATPES and MACE-OMOL/MH models are under the Academic Software Licence (ASL), which is limited to academic non-commercial use.
Where users get stuck
- Running MACE inside LAMMPS on GPUs (Kokkos, ML-IAP, multi-GPU, model compilation). #321#945#195#203
- Fine-tuning foundation models breaks across versions. #1031
- Dependency pins and data formats (e3nn pin, extended xyz parsing). #555#415
The choice it embodies: MIT code with a split model catalogue: early MPtrj-trained models are MIT, while newer models trained on larger datasets (OMat, OMol, MATPES) ship under the non-commercial ASL, so 'which checkpoint' decides whether commercial use is allowed.
Benchmark and reality: Matbench Discovery, where MACE is ranked, scores models as pre-filters against DFT-computed stability data, not experiment; a separate study found MACE-MP-0 (with M3GNet and CHGNet) systematically underpredicts energies and forces versus DFT on surfaces, defects, migration barriers and phonons, which it traces to near-equilibrium training data.
The question it answers: You ship MIT code but your best models are non-commercial: how do you decide per-model licensing, and how do you communicate it so industrial users don't adopt the wrong checkpoint?
Sources: raw.githubusercontent.com/…/HEAD/README.md · raw.githubusercontent.com/…/HEAD/ASL.md · nature.com/…/articles/s42256-025-01055-1 · nature.com/…/articles/s41524-024-01500-6
Meta FAIR Chemistry's repository of datasets, models and tooling for materials and molecular ML, centred on the UMA universal interatomic potential.
Weights: UMA weights on Hugging Face are manually gated: users must submit full legal name, date of birth, organization and country and accept the FAIR Chemistry License v1 (royalty-free, allows derivatives, but with an Acceptable Use Policy that bars e.g. military/nuclear uses, and requires acknowledgement in publications).
Where users get stuck
- Numerical problems in training and derivatives (NaNs, non-determinism). #1568#1618#563
- Calculator and checkpoint compatibility across versions (missing stress, loading older checkpoints, migrating to UMA). #888#936#1201
The choice it embodies: MIT code with weights behind a manual identity-verified gate and a custom license: Meta keeps the model 'open' in the sense of downloadable and modifiable, while adding use restrictions and a know-your-user step that an MIT checkpoint would not have.
Benchmark and reality: Benchmarks like Matbench Discovery score against DFT data rather than experiment, and fairchem's own README warns that UMA/OMat24 models are trained on DFT/DFT+U total energies not compatible with Materials Project calculations without corrections, so a leaderboard number does not transfer directly to a given workflow.
The question it answers: Should a research lab gate open weights behind an identity form and custom license? What adoption do you lose, and what risk or accountability do you gain?
Sources: huggingface.co/…/facebook/UMA · raw.githubusercontent.com/…/HEAD/README.md · nature.com/…/articles/s42256-025-01055-1
A Python library for representing and analysing crystal structures and molecules, with I/O for many electronic-structure codes; it powers the Materials Project.
Weights: Not a model; no weights.
Where users get stuck
- Parsing and handling VASP files (WAVECAR, WAVEDER, POTCAR hashes). #1585#219#1746
- Structure analysis edge cases: symmetry detection and neighbour finding. #3006#3888#1321
- API stability for downstream packages under calendar versioning. #2083#1985
The choice it embodies: A permissively licensed infrastructure library that ships frequent calendar-versioned releases (two on 2026-09-23 alone) while downstream projects ask for semantic versioning; its MP2020 energy corrections are also the reference other tools (MACE, fairchem) must document compatibility against.
The question it answers: You maintain the foundational library a whole field depends on: how do you balance release velocity against downstream stability, and what versioning contract do you offer?
Sources: github.com/…/materialsproject/pymatgen · github.com/…/pymatgen/releases · github.com/…/issues/2083 · raw.githubusercontent.com/…/HEAD/README.md
DeepModeling community, with core developers affiliated with DP Technology and AI for Science Institute, Beijing (AISI)LGPL-3.0-or-laterWeights · RestrictedLatest release v3.2.0 · 2026-08-192.1k stars
A package for training Deep Potential machine-learning interatomic potentials and running them in molecular dynamics (LAMMPS, i-PI, ASE and others), now centred on pretrained DPA4 models.
Weights: Pretrained DPA4-OMat24 checkpoints (Nano to Plus) are downloadable from Hugging Face without a gate, but under CC-BY-NC-4.0 (non-commercial).
Where users get stuck
- Silent correctness bugs: a maintained tracking issue lists bugs that give incorrect results without error messages. #2866#3270#5180
- Building and installing the C++/LAMMPS interface against TensorFlow and MD engines. #1809#2039#1957
- Training at scale: wall time, parallel training and memory. #2229#1774#4594
The choice it embodies: A community-branded LGPL project whose DeePMD-kit v3 paper lists DP Technology (Beijing) among author affiliations and whose DPA4 was announced jointly by AISI, Peking University, DP Technology and IAPCM; code is copyleft-lite, but the flagship pretrained weights are non-commercial.
Benchmark and reality: DPA4 is marketed as topping Matbench Discovery's combined score (as of 2026-05-22), but that benchmark ranks models as pre-screeners against DFT-computed stability, not experimental outcomes.
The question it answers: How does a company-backed 'community' open-source project stay credible as neutral infrastructure while its sponsor also sells commercial services, and why release code as LGPL but weights as non-commercial?
Sources: huggingface.co/…/deepmodelingcommunity/DPA4-OMat24 · arxiv.org/…/html/2502.19161v2 · blogs.deepmodeling.com/…//DPA4_05_22_2026 · nature.com/…/articles/s42256-025-01055-1 · Company profile →
Proteins and drugs
Structure prediction, co-folding and protein design. This group shows the open-versus-closed choice most clearly: the same class of model ships behind an application form, under MIT, or in size tiers.
The inference code for AlphaFold 3, which predicts the joint structure of complexes of proteins, nucleic acids, ligands and modified residues.
Weights: Since 23 July 2026 the parameters are a direct download, replacing an application form that Google DeepMind answered at its own discretion. The terms did not change: non-commercial organisations only, no use of the output to train similar structure-prediction models, and no redistribution of the parameters.
Where users get stuck
- GPU generations and memory: older cards (compute capability 7.x), consumer 24 GB cards and the newest Blackwell parts each drew their own issue, and a large complex ran out of memory even on an 80 GB H100. #59#209#341#394
- Chemistry the input format does not make easy: covalent ligands, leaving atoms and bonds to specific residues. #159#428
- Runtime at scale: large jobs take longer than the published timings suggest. #609
The choice it embodies: Access opened in three steps in 2026 without the weights terms moving: the code went from CC BY-NC-SA 4.0 to Apache 2.0 on 9 June (the commit says only the code was relicensed), the application form became a direct download on 23 July, and on 21 August the README began pointing commercial users to Google Cloud. Open code and a free download for academics sit beside a paid route for companies.
The question it answers: When is it right to open code and weights to academics while keeping commercial use behind a cloud API, and what does a no-training-on-outputs clause protect?
Sources: github.com/…/main/WEIGHTS_TERMS_OF_USE.md · google-deepmind/alphafold3 @ 7b197fe · google-deepmind/alphafold3 @ dd1a7ba · google-deepmind/alphafold3 @ 9b30e9a · Company profile →
Jeremy Wohlwend, Gabriele Corso, Saro Passaro and contributorsMITWeights · OpenLatest release v2.2.1 · 2025-09-084.2k stars
Boltz-1 and Boltz-2, open biomolecular structure models; Boltz-2 also predicts protein-ligand binding affinity.
Weights: Code and weights are both MIT, which the README states means free for academic and commercial use.
Where users get stuck
- How much GPU a job needs: memory against protein length, multi-GPU nodes, and whether it runs on a personal computer at all. #8#17#15#341
- Kernels and packages: the triton dependency and custom kernels that slowed things down on an RTX 4090. #433#391
- Using it as a screening tool: many ligands against one protein, and templates. #570#334
The choice it embodies: Fully permissive from the first release: MIT for code and weights since Boltz-1 in November 2024, positioned in its own README as the first fully open model to approach AlphaFold 3 accuracy. Openness is the product's differentiation, not a concession.
Benchmark and reality: The README says Boltz-2 is the first deep-learning model to approach the accuracy of physics-based FEP while running 1000x faster. The same README has said evaluation scripts and affinity predictions on the FEP+ benchmark are 'coming soon' since the Boltz-2 release in June 2025.
The question it answers: If your model is MIT-licensed and a competitor can run it for free, where does the business come from, and what do you owe users when a headline claim's evaluation code has not shipped?
Sources: raw.githubusercontent.com/…/HEAD/README.md · github.com/…/main/LICENSE · doi.org/…/10.1101/2025.06.14.659707
Chai-1, a multimodal model for predicting the structure of proteins, nucleic acids, small molecules and covalent modifications, runnable without MSAs.
Weights: Chai-1 code and weights are Apache 2.0, stated in the README as usable for commercial purposes including drug discovery. Chai-2, the antibody-design model, is described in a technical report and is not in this repository.
Where users get stuck
- Hardware: users asking how to get an A100, whether it runs on CPU, and hitting bfloat16 kernel errors on older GPUs. #69#19#26
- Size limits and cluster use: requests for crops beyond 2048 tokens, pre-downloading weights, and running on a shared cluster. #153#131#318
- The same input giving different results on the web server and a local install, still open. #155
The choice it embodies: Chai-1 launched in September 2024 under a non-commercial community licence and moved code and weights to Apache 2.0 two and a half months later. The next model, Chai-2, stayed out of the repository, and the README's route to it is a partnerships email.
Benchmark and reality: The README notes that the results in the preprint and on the web server use a different MSA search than the public MMseqs2 server the open package calls, so a local run is not the configuration the reported numbers came from.
The question it answers: Open the last-generation model, keep the next one closed: when does releasing yesterday's model build the trust and pipeline that sells tomorrow's?
Sources: raw.githubusercontent.com/…/HEAD/README.md · chaidiscovery/chai-lab @ 0c7cfd5
Code for the ESM3 generative protein language model and the ESM C embedding models; the repository formerly at evolutionaryscale/esm now redirects here.
Weights: Tiered by size: the 1.4B esm3-sm-open-v1 weights are on Hugging Face under MIT, while the 7B and 98B flagship ESM3 models are reached through Biohub's API. Use is also subject to Biohub's acceptable-use policy.
Where users get stuck
- Reproducing the paper: an issue open since July 2024 reports LDDT on CAMEO and CASP14 that does not match the published numbers. #63
- Getting embeddings out and loading models: ESM C embeddings, local model paths, and tokenizer mix-ups between ESM C and ESM3. #176#132#155
- Running the quickstart at all, on Colab or after download. #89#5#285
The choice it embodies: Until May 2026 the repository ran four licences at once: an open licence for code and ESM C 300M, a non-commercial one for the ESM3 open model and ESM C 600M, free non-commercial API credits, and a paid commercial API. On 27 May 2026 it was relicensed to MIT under Chan Zuckerberg Biohub copyright. The size tiers stayed: the flagship models are still API-only.
The question it answers: How do you tier a model family between open weights and an API, and what happens to a licensing ladder when the organisation behind it changes hands?
Sources: github.com/…/_assets/ESM3_README.md · Biohub/esm @ c94ed8d
Structure prediction in a Colab notebook or a local batch tool, with MSAs generated by a public MMseqs2 server instead of a local database search.
Weights: ColabFold does not ship a model of its own; it runs AlphaFold2 and related models, and its contribution is the MSA server and packaging. Chai-1 and ColabFold's own AlphaFold 3 input export call the same server.
Where users get stuck
- The shared MSA server: connection failures, timeouts under batch load, a Forbidden error, and a batch run that still called the server when MSAs were supplied. #339#481#606#664#764#563
- Local installs: CUDA versions for the Docker image, memory errors during alignment expansion, and running out of GPU memory on complexes. #549#691#8
The choice it embodies: It took the two costs that kept structure prediction in specialist labs, a GPU and a large sequence-database search, and moved them to free Colab GPUs and one shared server. The README's own FAQ asks users to query that server serially from a single IP and notes a self-hosted search needs about 128 GB of RAM, so the barrier it removed is now a shared resource others depend on.
The question it answers: Adoption came from removing setup, not from a better model: who pays for the shared server everyone now depends on, and what is the product when the model is someone else's?
Sources: raw.githubusercontent.com/…/HEAD/README.md
AI scientists and agents
Systems that hand literature search, experiment design and writing to an LLM. There are no weights to discuss; what to read is whose model it depends on, and how much of its output has been checked by anyone else.
PaperQA2: a Python package and CLI (`pqa`) for agentic retrieval-augmented generation over PDFs, text, Office documents and source code, focused on answering questions from scientific literature with citations.
Weights: No model weights are released; PaperQA2 calls external LLMs and embedding models through LiteLLM, defaulting to OpenAI models and embeddings, with llamafile/llama.cpp, Ollama and Sentence Transformers documented as local options.
Where users get stuck
- Pointing it at a local or non-OpenAI model: users configure vLLM, Ollama, Gemini, Bedrock or an OpenAI-compatible endpoint, but indexing or querying still tries to reach OpenAI or asks for OPENAI_API_KEY. This keeps coming up from 2024 into 2026. #582#789#957#1321#667
- PDF ingestion failures: encoding errors on some papers and image formats that the multimodal enrichment step rejects. #1068#1273
- Installation and dependency conflicts, for example on Google Colab. #1238#443
The choice it embodies: The open library ships the RAG and agent loop, but you bring your own PDFs and your own LLM keys. The README FAQ says FutureHouse's internal version uses different tools plus paper-access APIs and licenses it cannot share. FutureHouse's hosted platform (May 2025) offers a literature agent, Crow, through a free web UI and API with a built-in paper corpus. Edison Scientific, a for-profit FutureHouse spinout from late 2025, now handles commercialization.
Benchmark and reality: The README links the PaperQA2 paper's 'superhuman' results, but its own FAQ ('How come I get different results than your papers?') says those results used internal tools and licensed paper access. It also says the paper's agent had to find relevant papers across the whole literature rather than starting from the known relevant PDFs, so open-source users should not expect to reproduce the published numbers as-is.
The question it answers: You open-source the core of your agent but keep the data access (licensed full-text corpora) that produced your headline benchmark. How do you position the open library against your hosted product, and what do you tell users whose results don't match the paper?
Sources: raw.githubusercontent.com/…/HEAD/README.md · futurehouse.org/…/research-announcements/launching-futurehouse-platform-ai-agents · techfundingnews.com/…//edison-raises-70m-ai-scientist-platform · arxiv.org/…/abs/2409.13740 · Company profile →
The AI Scientist (v1) is a pipeline in which LLMs generate research ideas, edit and run experiment code from human-written templates (NanoGPT, 2D Diffusion, Grokking), write a LaTeX paper and review it automatically.
Weights: No model weights; it runs on API models (OpenAI, Anthropic, DeepSeek, OpenRouter and others). The code license (Dec 2025, replacing Apache 2.0) is a Responsible-AI-style license that requires prominent disclosure that any manuscript was machine-generated.
Where users get stuck
- Literature search through Semantic Scholar: users hit 429 rate limits without a key and wait for S2 API key approval. #84#104#78
- Running without paid APIs or without a Linux/NVIDIA machine: requests for local models via Ollama, Apple Silicon, and non-developer setup went unanswered. #92#43#124
- Pushback on the framing itself: an open issue argues that 'AI scientist' overstates what the system does. #198
The choice it embodies: The code is open and cheap to run (README: 'Typically less than $15 per paper with Claude Sonnet 3.5'), but it depends on frontier APIs and it executes LLM-written code. The README tells users to containerize and restrict web access, and the only Docker image is community-contributed. The 2025 license change adds a legally binding AI-disclosure duty instead of relying on norms.
Benchmark and reality: An independent evaluation (Beel, Kan & Baumgart, arXiv 2502.14297) found poor novelty assessments and 42% of experiments failing from coding errors. It also found papers with a median of five citations, frequent structural errors, some hallucinated numerical results, and quality it likened to 'a rushed undergraduate paper'. Still, each paper cost USD 6 to 15 with about 3.5 hours of human involvement.
The question it answers: Your open-source agent autonomously runs code it wrote and produces papers indistinguishable at a glance from human work. What guardrails do you ship in the product (sandboxing, licensing, disclosure), and which do you leave to users?
Sources: raw.githubusercontent.com/…/HEAD/LICENSE · raw.githubusercontent.com/…/HEAD/README.md · arxiv.org/…/abs/2502.14297v3
The AI Scientist-v2 is a template-free successor to v1 that uses agentic tree search, guided by an experiment-manager agent, to generate hypotheses, run ML experiments and write papers. Sakana says it produced the first fully AI-written paper accepted through peer review, at an ICLR 2025 workshop.
Weights: No model weights; it runs on OpenAI, Gemini and Claude-via-Bedrock APIs. It uses the same AI Scientist Source Code License as v1, with mandatory disclosure of machine-generated manuscripts.
Where users get stuck
- Runs stall or fail on the API side: endless backoff in the initial implementation stage, request timeouts, and credential errors (for example AWS/Bedrock sessions). #34#50#4#14
- Pipeline bugs in the write-up and summary stages, plus requests to use models outside the built-in list. Ollama support was merged and then reverted in October 2025. #67#42#6
The choice it embodies: v2 gives up v1's human-written templates for open-ended exploration. Its own README says v2 'doesn't necessarily produce better papers than v1', because v1 has higher success rates when a strong template exists and v2 has lower success rates. It costs about $15 to $20 per run for experiments with Claude 3.5 Sonnet, plus about $5 for writing. Like v1, it must be sandboxed because it executes LLM-written code.
Benchmark and reality: Sakana submitted 3 AI-generated papers to the ICLR 2025 ICBINB workshop. One scored 6/7/6 (average 6.33, above the workshop's average acceptance threshold) and was withdrawn before publication as planned. Humans picked which papers to submit, and Sakana said none of the three met its own bar for an ICLR main-track paper, noted 'embarrassing citation errors', and pointed out that workshop acceptance rates (60-70%) are far higher than the main conference's (20-30%).
The question it answers: How do you announce a milestone like 'first AI paper to pass peer review' so it is accurate? Which caveats (human selection, workshop vs main track, withdrawal, citation errors) go in the headline, and which go in the fine print?
Sources: raw.githubusercontent.com/…/HEAD/LICENSE · raw.githubusercontent.com/…/HEAD/README.md · sakana.ai/…//ai-scientist-first-publication
