LLM Benchmarks for Pharma: State-of-the-Art Results

👉 Efficiency + privacy with SOTA performance. In partnership with Insilico Medicine, we adapted a compact LFM2 model for drug from Liquid AI for discovery workloads — and it’s delivering strong, state-of-the-art results across key benchmarks. ⚡ Small footprint, high performance ⚡ Competitive with much larger models on ADMET, retrosynthesis, and molecular reasoning ⚡ Runs efficiently on CPUs/NPUs/GPUs What stands out to me isn’t just the benchmarks — it’s the combination: → Efficiency: usable without massive infrastructure → Privacy: sensitive molecular data can stay local, no external calls required → Performance: still hitting SOTA on targeted tasks That combination is what unlocks real adoption in domains like pharma. Feels like a glimpse of where applied AI is heading.

If you like high intelligence density, efficiency and elegance, you got to try Liquid AI foundation models from the genius Ramin Hasani and team. After #MMAIGym, these models become superintelligent in many drug discovery tasks and you can run them on your machine, hook them up to your OpenClaw, or put them into your playground.

⚡ 99% of the world’s computers exists outside of data centers. Yet, most #drugdiscovery teams are still tethered to massive, high-latency flagship models that frequently fail to target performance goals for #ADMET and toxicity. Insilico Medicine and Liquid AI are changing that. We’ve partnered to train LFM2-2.6B-MMAI, a model designed for the edge, but better everywhere. After a single training session in the #MMAIGym, this "tiny" 2.6B model is now outperforming models like TxGemma-27B while maintaining a significantly smaller energy footprint. Key Benchmarks at a Glance: -ADMET Supremacy: Surpassed TxGemma-27B across multiple property prediction endpoints. -94% #Retrosynthesis Accuracy: Achieved via specialized SFT+RFT training sessions. -Undisputed SOTA: Leading performance in functional group reasoning (FGBench) and molecular optimization (MuMO). -Edge-Ready: Optimized for CPUs, NPUs, and GPUs, providing low-latency, "always-on" intelligence without the need for cloud calls. 🤝 We’re ready to help you integrate these models directly into your DMTA cycle for real-time lead optimization. See it in action: Test the MMAI Demo today [https://lnkd.in/esyGn6g9]

Biology is entering a new era—one defined by inference at scale. From predicting complex biomolecular interactions to accelerating therapeutic discovery, advancements like OpenFold3 are reshaping what’s possible. With high-performance compute at the core, AMD Instinct GPUs are helping bring these models into real-world use—delivering the speed, memory capacity, and scalability needed to move from isolated predictions to continuous, production-grade discovery. AMD #AMDBrandAmbassador

For whoever needs to hear this. Self hosted AI inference is converging on memory bandwidth and throughput as the bottleneck more so than total model hosting memory capacity. Dense models with extended thinking are making it possible to get a lot of performance out of a smaller total memory envelope than ever. But despite having lower total parameters than similarly performing models of the past, they are exceedingly token heavy due to their reasoning budgets. TLDR: Don’t overestimate what you can do with big models on slower unified memory and don’t underestimate what you can do with smaller dense models on dedicated GPUs.

Second international acceptance of the week. Excited to share that my Bio × GPU paper has been accepted for a short talk at Arch4Health, ACM ICS 2026 (UK 🇬🇧). Paper: "From 805 ms to 23 ms: Accelerating State-Space Models for Real-Time ICU Monitoring with Fused Kernels" Real-time ICU early-warning systems operate with sub-50 ms latency to enable timely clinical intervention. But irregular sampling and >30% missing values in physiological time series force deep learning pipelines to use sequential preprocessing, which consumes over 85% of wall-clock time. We introduce accelerated State-Space Models with a time-aware formulation that fuses interpolation and inference into a single kernel, eliminating this bottleneck. Results: → 35.7× lower end-to-end latency (805 ms → 23 ms) vs Optimised PyTorch SSM → 2.5× faster training than GRU-D baselines → 10,316 samples/sec throughput → Hardware-portable across NVIDIA and AMD GPUs →Real-time multi-patient ICU monitoring on a single consumer-level GPU. Grateful to Embedded LLM for the trust and GPU resources to push Bio × GPU systems work into places where milliseconds matter for patients 🙌 #ICS2026 #Arch4Health #HPC #GPUComputing #Triton #AI4Health #BioAI #SystemsML

Sponsored: Advances in embedded processors have opened the field for bringing increased AI intelligence to the edge. However, one major challenge has emerged: Traditional edge AI solutions (graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and application-specific integrated circuits (ASICs)) have bounded applicability because most of them are limited in terms of flexibility. Typically, ASICs and/or high-power-consumption GPUs and FPGAs fall into this category. Looking for answers to the shortcomings of existing products, engineers and designers examined the use of embedded processors, for both lower-end, resource-constrained consumer products as well as complex industrial operations. The solution? The TinyEngine neural processing unit (NPU). Read more here: https://bit.ly/48JS2EA

Pretraining benchmarks often require massive multi-node infrastructure - which can create a barrier to entry for benchmarking. For MLPerf Training v6.0, the task force (led by AMD, NVIDIA, and NIT University) is introducing GPT-OSS 20B: a new, highly accessible Mixture-of-Experts (MoE) pretraining benchmark. With 21B total parameters but only 3.6B active per token, GPT-OSS 20B allows researchers and organizations to evaluate the complex routing logic of MoE architectures from scratch on configurations as small as a single 8-GPU node. Because MoE models trained from randomized weights are notoriously noisy, the task force implemented three critical interventions to drive the Coefficient of Variation (CV) down from 15% to under 5%: 1️⃣ A static, unshuffled evaluation set to eliminate routing jitter 2️⃣ Optimizer stabilization (ϵ=10−5\epsilon = 10^{-5}) 3️⃣ Strictly defined weight initialization standards Read the technical breakdown of how the task force engineered stability into MoE benchmarking: https://bit.ly/3QLwvVU #MLPerf #GPTOSS #MoE #AIHardware #MachineLearning #MLCommons

https://lnkd.in/g-XvX89a 🚀 AIBIZ’s semiconductor AI case study is now featured by AMD. Together with AMD and HPE, AIBIZ demonstrated how lightweight industrial AI can deliver real-world impact in semiconductor manufacturing without relying on GPU-heavy infrastructure. Highlights from the case study: • 30% faster AI inference with AMD EPYC CPUs • Real-time wafer defect prediction • Improved semiconductor yield and reduced production loss • Scalable deployment across global fabs Proud to see AIBIZ’s technology and vision recognized on a global stage. 🌍 #AMD #CaseStudy #AMDPartnership #AIBIZ #ManufacturingAI

𝐓𝐞𝐜𝐡𝐧𝐢𝐜𝐚𝐥 𝗔𝗿𝘁𝗶𝗰𝗹𝗲 🔹EPAC: A #RISCV Accelerator from the European Processor Initiative https://lnkd.in/d3egcxFe This paper presents EPAC, a RISC-V-based #accelerator chip developed within the European Processor Initiative (EPI) as part of a multi-year, multi-partner effort to build a European HPC processor ecosystem. EPAC is implemented in GlobalFoundries 22FDX (#GF22FDX) technology, covers an area of 27 sq mm with approximately 0.3 billion transistors, and integrates three distinct RISC-V compute tiles targeting different workload classes: VEC, a vector processing tile for double-precision #HPC workloads; STX, a many-core tile optimized for stencil and machine learning computations; and VRP, a variable-precision tile for iterative numerical solvers requiring extended floating-point formats. Read more at https://lnkd.in/d3egcxFe #semiIP #IPcores #semiconductor