Alibaba Open Sources LOGOS: 1B Parameters Beats Microsoft’s NatureLM That’s 56× Larger — The Era of Unified Scientific Grammar Has Arrived

On June 18, Alibaba ATH-Token Foundry, together with the Gaoling School of Artificial Intelligence at Renmin University of China, dropped a heavyweight release — LOGOS (Language Of Generative Objects in Science) — claimed to be the first multi-domain scientific foundational generative model based on a unified “scientific grammar.” The model is now open source, and with only 1B parameters, it consistently matched or outperformed Microsoft’s massive 8×7B NatureLM on six representative scientific tasks.

A win across domains with only 1/56 of the parameters — the number itself is eye-catching, but what’s more noteworthy is the underlying idea: packing proteins, small molecules, materials, complexes — these “speaking different languages” scientific objects — into the same token space, letting a decoder-only autoregressive model do next-token prediction. Sounds simple, but getting this to work is actually somewhat counterintuitive.

First, let’s clear up the comparison

Microsoft’s NatureLM also follows the sequence-based model route. When it was released last year, it was arguably the first explicit “natural language” player in AI for Science — treating small molecules, proteins, materials, DNA, RNA all as sequences. Its architecture uses Mixtral-style MoE, 8×7B, with a sizable amount of active parameters.

LOGOS went straight in with a dense 1B model. In Alibaba’s evaluation, across six tasks it either matched or exceeded NatureLM. This isn’t to say NatureLM is bad, but rather that when grammar design is truly unified, parameter efficiency can make a remarkable leap. NatureLM essentially stitched together tokenizers from different domains and relied on huge parameters to brute-force handle cross-domain differences; LOGOS instead “translates” these heterogeneous objects into the same language at the vocabulary level, so the model doesn’t have to spend parameters doing implicit translation.

This approach isn’t unfamiliar in the LLM community — GPT once unified everything into text-to-text tasks with a similar logic. But in the scientific domain, fully tokenizing 3D structures, chemical reaction diagrams, crystal periodicity is far trickier than in NLP.

The 44.87B-token Pretraining Corpus — How it’s built

LOGOS’ pretraining corpus is quite interesting, covering 7 modality categories, organized hierarchically:

• Biomacromolecule layer: proteins 28.9B tokens + antibodies 3.0B tokens
• Chemical entities & transformation layer: small molecules 2.1B tokens + chemical reactions & MOF materials 0.47B tokens
• Interface interaction layer: protein pockets 5.8B tokens + protein-pocket–ligand complexes 4.6B tokens

Looking at this distribution, proteins dominate, which aligns with current AI for Science data availability — UniProt, PDB, those databases have plenty of volume. Small molecules and materials have fewer tokens, but chemistry is inherently dense — 2B tokens of SMILES/SELFIES already cover a significant portion of chemical space.

The standout design is in the “interface interaction layer.” Protein pockets and protein–ligand complexes total 10.4B tokens, nearly one-quarter of all data. This means LOGOS isn’t just stacking single sequences for pretraining, but is explicitly feeding “interactions” as an independent modality.

How the “text description method” eliminates 3D geometric networks

This is LOGOS’ most technically meaningful part — also the one I think is most worth digging into.

Traditionally, predicting protein–ligand binding forces you onto SE(3)-equivariant neural networks, geometric Transformers, coordinate regression, etc. E3NN, EquiformerV2, Uni-Mol — all follow that route. The reason is simple: in 3D space, molecular–protein interactions depend heavily on atomic coordinates, and rotational/translation invariance is a hard constraint.

LOGOS takes a different route: directly “grammatizing” 3D spatial contact patterns into discrete tokens. They don’t give detailed references, but from the description, it seems they discretize the contact relationships (distance, angle, type of interaction) between pocket residues and ligand atoms into symbolic sequences, then train the model to learn these spatial rules purely via sequence prediction.

The advantages are obvious:

1. No need to input 3D coordinates, making inference lightweight
2. No need to maintain equivariant networks, drastically reducing engineering complexity
3. Seamless integration with text/sequence tasks, no backbone switching

The drawbacks are predictable: information loss. Compressing continuous spatial relationships into discrete tokens will inevitably lose details. But LOGOS’ experimental results show that as long as the vocabulary is well designed and the training data is sufficient, a sequence model can “fill in” the 3D interaction patterns. This is similar to how LLMs can perform decent physical reasoning without an actual physics engine.

In essence, this is trading scale and representation for inductive bias. Deep learning’s past 20 years have repeatedly shown that, given enough data and compute, inductive bias can often be replaced by more general architectures.

Form–Objective Alignment: Solving scientific AI’s most awkward issue

Anyone who’s worked on deploying AI for Science knows the pain point: Pretraining objectives and downstream tasks often misalign severely.

If you pretrain a molecular encoder with contrastive learning, and your downstream task is generation, you need to attach a decoder and train from scratch. If you pretrain protein representations with masked language modeling, and then need to predict structure, you have to switch architectures. Every transition between research steps means starting over with model, objective, and assumptions. AlphaFold2 can’t be used for docking; ESM can’t be used for retrosynthesis — that’s normal.

LOGOS tackles this with two alignments:

• Form alignment: the sequence format of pretraining data = input/output format for downstream tasks
• Objective alignment: pretraining’s next-token prediction = downstream conditional generation objective

This is what they call form–objective alignment. In other words, once pretrained, any downstream task becomes “give a prefix, then complete the sequence,” with no need for complex adapter layers or architecture retraining.

Essentially, this is bringing the GPT paradigm into the scientific field. OpenAI showed with GPT-3 that “large enough language models + prompts” can replace many specialized NLP models; Alibaba is aiming to prove that “well-designed scientific grammar + autoregression” can replace many specialized scientific models.

1B vs 8×7B: Why such a huge gap

Parameter efficiency doesn’t appear out of nowhere — I tend to attribute it to three factors:

First, unified vocabulary enabling parameter reuse. NatureLM, though unified as a model, doesn’t fully share tokens across domains — the model needs to maintain multiple “mental models” for varying modalities. LOGOS’ shared vocabulary processes protein tokens and small molecule tokens in the same embedding space, boosting parameter reuse.

Second, explicit modeling of interaction modalities. The 10.4B tokens of pocket and complex data effectively feed the model a “dictionary” of how molecules and proteins interact. NatureLM largely learns this implicitly from single-domain data and some cross-domain generation tasks — far less efficient.

Third, dense architecture vs. MoE architecture in small-scale settings. MoE works well at large parameter scales, but in small-scale models the routing overhead can be a burden. At 1B, dense models get better active parameter utilization.

Of course, the 1/56 figure is a bit of a simplification — NatureLM’s actual active parameters are ~13B, not 56B. But even by active parameters, 1B vs. 13B is a 1/13 gap — still very significant.

What this means for developers

If you work on drug discovery, materials design, synthesis planning, LOGOS has at least three points worth trying:

1. Reuse the same model weights across tasks — no need to maintain separate checkpoints for pocket prediction, molecule generation, retrosynthesis; inference infrastructure can be greatly simplified.
2. 1B scale is runnable — a single RTX 4090 can handle it, making local lab deployment feasible.
3. Prompt-based task definition — “give prefix, complete” means researchers can quickly try new tasks without writing training code.

Weaknesses should be noted: reference materials don’t disclose full benchmark details — exactly which metrics matched or exceeded NatureLM, how it compares to the latest domain SOTAs (e.g., RFdiffusion, Boltz, Chai-1) — currently unseen. Once the open-source model weights and evaluation pipeline are released, the community will quickly verify.

Another open question is scaling behavior. LOGOS-1B is just a start — if the scientific grammar design really holds, LOGOS-7B and LOGOS-70B should show a steeper scaling curve than NatureLM. This is worth watching over the next six months.

Final thoughts

For the past two years, AI for Science’s main theme has been “bigger is better” — AlphaFold series, RoseTTAFold, Boltz, various diffusion-based docking models — everyone’s stacking parameters, data, geometric priors. LOGOS offers another path: unify representations first, then let the model learn. This mirrors NLP’s path from word2vec to GPT — moving from “design features for each task” to “let a general model learn general representations.”

Microsoft’s NatureLM started the journey; Alibaba’s LOGOS goes deeper on the same path with smaller parameters. For domestic AI-for-Science teams, having access to a 1B-scale open-source foundation that covers protein–small molecule–materials lowers the barrier by more than a notch. Model weights and code are already open source — those interested can pull and run it directly.

---

References

• Alibaba open sources its first unified scientific model LOGOS, surpassing Microsoft’s NatureLM with only 1/56 of the parameters - IT Home: Main Chinese report on LOGOS, including model architecture, pretraining corpus composition, and form–objective alignment design details.