DeepSeek 4.1 Flash dropped last Friday and we just completed a full benchmarking pass with the model on our harness. The early results look pretty incredible with DeepSeek performing at near Opus 5 levels at about 1/12th the cost on DrugDiscoveryBench tasks by @scale_AI . For certain tasks like literature-based compound and biochemical analysis we're even seeing DeepSeek outperform Opus significantly. We'll be continuing to optimize it on our harness but really excited about the implications on cost and search space exploration for drug discovery utilizing the cheaper open weight models. We've been working with @baseten for our open weight model inference and them having the model ready on day 0 really helps us get results out quickly!

Our first event was so good, we're hosting another one in Boston! We'll be talking about frontier agent development for drug discovery, so come on by 🪩 luma.com/rq9eyuox?tk=zWEzCY

Very cool writeup by @mirrorphysics about how they are using @modal for drug discovery. We're seeing a ton of growth in computational bio. mirrorphysics.substack.com/p… Who knows, maybe one day I'll be taking drugs that were found on a Modal GPU.

To truly unlock scientists in drug discovery, the tools they rely on need to scale safely, reliably, efficiently, and dynamically. Our AI platform Axon hosts state-of-the-art tooling across preclinical drug discovery and is powered by Modal's serverless cloud infrastructure, allowing these tools to be deployed seamlessly across workloads of any size. For users, the result is access to industry-leading capabilities, without the infrastructure burden: 🪩 Compute costs always go towards advancing discovery 🪩 Parallel scaled execution expedites progress 🪩 Fast custom integrations for bespoke workflows 🪩 Comprehensive industry-leading security Details in our newest blog mirrorphysics.substack.com/p…

If AI models can meaningfully improve that filtering process for both generating proteins and evaluating those worth testing, the impact could be measured in fewer experiments, lower costs, and faster paths to new drug therapies. Exciting to see portfolio companies @boltz_bio and @mirrorphysics collaborating to build the new stack for accelerating drug discovery with AI

Huge release from the Boltz team. The API is especially exciting: frontier biomolecular models becoming agent-native primitives. We’re excited to partner with @boltz_bio to bring these capabilities directly into our preclinical discovery agent, Axon.

It’s been a busy time at Boltz already! Today, we are announcing: - BoltzMol-1 - BoltzProt-1 - Boltz API Delighted to have fantastic partners @benchling, @phylo_bio, @amazon Bio Discovery, @RowanSci , @tamarindbio, Kiin Bio, @PaulingAi, @mirrorphysics and @CultivariumFRX

⚛️ EquiformerV3 is here ! State-of-the-art on Matbench Discovery and OC20🏆 Better scaling, improved sample efficiency, stronger generalization. The Equiformer line keeps getting better! Huge congratulations to Yi-Lun Liao and Tess Schmidt for driving this vision so brilliantly ! It's been inspiring to watch it come to life at this level. Having witnessed & used previous Equiformer versions, I am grateful to have contributed to this one, even if it was a small piece of a much larger puzzle. Go check it out 🤗

🏆🏆🏆 EquiformerV3 just topped MatBench Discovery using less than 1/3 of the compute of the closest competitor 🏆🏆🏆 EquiformerV3 precisely simulates chemical physics by scaling SE(3)-equivariant graph attention transformers. AND it was released on @huggingface 🤗

EquiformerV3 + DeNS is currently the best method on Matbench Discovery (as of April 14th, 2026). Paper: arxiv.org/abs/2604.09130 Code (training + eval): github.com/atomicarchitects/… (We are standing on the shoulders of giants to push the frontier of AI for atomistic simulation. 🙏 More to come. 🙂

Today, the team at Mirror is delighted to announce EquiformerV3: a new AI model for the precise simulation of chemical physics, made in collaboration with @yilunliao from the Atomic Architects group of Professor @tesssmidt at @MIT At time of writing, EquiformerV3 tops Matbench Discovery, one of the most widely-used benchmarks in computational materials science, while using less than a third of the training compute of the closest competitor. We will release additional checkpoints of EquiformerV3 in the coming weeks, so stay tuned. Check out the current release + more details below! Announcement: mirrorphysics.substack.com/p… Code: github.com/atomicarchitects/… Paper: arxiv.org/abs/2604.09130v1

At the Bone Health Research Group at @imperialcollege, researchers James Rowe and Richie Abel look to the forefront of technology to help uncover treatments for pathologies like Osteogenesis imperfecta (O.I.), or brittle bone disease. This rare disease arises from a complex cluster of genetic mutations that damage the mechanical integrity of collagen, the body’s essential structural protein. Recently, James used one of Mirror’s models to capture the rupture behavior of a collagen fragment. A remarkable detail emerged: under stress, each strand is predicted to break at a proline backbone C-C alpha bond, right next to glycine residues where the most problematic O.I. mutations occur. This provides a glimpse into the chemical mechanisms behind O.I., helping build the case for clinical targets. This is just the beginning. With the help of experts like James and Richie, accurate and transferable simulation can tackle complex challenges in many fields.

Thankful to partner with @nvidia and @huggingface as a first user of their new Training Cluster as a Service platform, applying DGX Cloud Lepton to produce high-fidelity chemical models at large scale. Exciting results to come.