Accepted to the #EMNLP2026 Industry Track 🎉 @InSilicoMeds built MMAI Gym for Science. @liquidai built the Liquid Foundation Model. We combined them. 400+ drug-discovery tasks. Chemistry-native tokens. SFT, then RL. Not a bigger model. A better gym. 🏋️ #insilicoSOTAFM 🧵👇

5/6 The same model can also write 3D molecular structures as text 🧬 Nothing has to condition it: asked for a molecule and nothing else, the model writes the graph and every atom’s coordinates. Given a protein pocket, it does the same thing on demand - the ligand’s SMILES graph first, then coordinates one atom at a time. No diffusion module. We measure this on #Bench3DFit, our own benchmark for pocket-conditioned generation: both LFM2-MMAI variants achieve the strongest UniDock binding affinity of any language model we tested - beating several specialist diffusion baselines and approaching the strongest models overall. The contrast is striking: GPT-5.5 generates molecules that are almost perfectly valid and geometrically tidy - but bind poorly. Our models do the opposite. MMAI Gym pushes the LM toward the objective that matters for structure-based drug design, rather than rewarding it for simply playing safe 🎯

LFM2-2.6B proved remarkably efficient for its size, providing a strong foundation for training a SOTA-level specialist on our tasks. Small model, serious chemistry.

🪶 What can a 2.6B-parameter language model achieve in retrosynthesis? With a leak-proof, broad-coverage drug discovery benchmark #DDDBench, it becomes easier to evolve general-purpose models into SOTA-level specialists. #insilicoSOTAFM

We’ve updated the #DDDBench with our own models, including LFM2-MMAI-Chem-SSRS-2.0, our single-step retrosynthesis specialist and the current SOTA on #ChemCensor.

I’m excited to share the follow-up paper to our #ChemCensor framework for evaluating single-step retrosynthesis 🧪 Our C3LM model was already competitive with, and in some cases superior to, general-purpose foundation models, but a gap remained between LLMs and conventional retrosynthesis models. We finally closed and surpassed that gap 🚀 The key ingredients were… 👇

3/n Together, these two ingredients enable C3LM to compete with and surpass strong conventional, non-LLM retrosynthesis models on the challenging URSA-expert-2026 test set 🏆 Our analysis shows that C3LM and conventional retrosynthesis models explore complementary reaction spaces 🧩 Each approach generates plausible reactions that the other may miss, suggesting that hybrid ensembles could provide broader chemical coverage and more robust synthesis planning than either model family alone 🤝

Huge progress from @AnthropicAI and a fascinating case study with @adaptyvbio. Protein binders are a reasonable testbed for AI in drug discovery: compared with many other modalities, they’re relatively fast and inexpensive to synthesize, screen, and validate experimentally. This is an important milestone, but not the finish line. Our own benchmarks show that Opus is steadily improving on drug-discovery tasks, yet it still struggles to generate physically plausible small-molecule ligands, especially relative to specialized diffusion models. There remains a long road from designing protein binders to developing safe and effective drugs.