AI Models
Model releases, benchmarks, inference and fine-tuning.
LFM2.5-24B-A2B on MLX-Serve: Mixed 4/8-Bit Conversion and Benchmarks
A focused guide to the public LFM2.5 mixed-precision conversion for local AI deployment, its tensor choices, and recorded results.
Read Article
From Climate Concern to Climate Action: Countering Misinformation and Supporting Behavior Change
How climate misinformation can blunt accurate communication, and what research on psychological inoculation suggests for helping people evaluate claims and act on climate change.
When Machines Outguess Neuroscientists About Their Own Research
UCL's BrainBench benchmark tested 15 LLMs against 171 neuroscience experts at predicting study outcomes. The models hit 81.4% accuracy versus 63.4% for humans — and a neuroscience-specialized BrainGPT reached 86%.
Navigating FACET Model Weights and Protein Backbone Torsion Prediction
Explore the model weights, retrieval indices, and benchmark performance behind FACET, an open-source tool predicting protein backbone torsion angles from NMR chemical shifts.
Nvidia's KVTC Breakthrough Cuts LLM Memory Footprint by 20x Without Weight Changes
A detailed look at Nvidia’s KV Cache Transform Coding (KVTC), examining how it compresses conversation history up to 20x without altering model weights, its impact on time-to-first-token latency, and integration plans for inference engines.
Flat Budgets, AI Ambitions: What 500 CISOs Actually Bought in 2026
Security budgets grew a whopping 5% in 2026 — yet AI dominated net-new spending. IANS senior director Nick Kakolowski and the firm's joint benchmark with Artico Search break down the paradox, the vendor pricing squeeze, where AI dollars actually land, and why nobody's getting fired.
Artificial Intelligence Environmental Cost: How Memory Bandwidth Shapes LLM Inference Latency
See how model size, memory bandwidth, and network latency shape LLM inference speed, serving costs, and the environmental footprint of AI compute.
The AI Cost Curve and Comprehensive LLM Cost Breakdown: Lower Token Prices, Bigger Enterprise Decisions
Research notes for an article on why falling model prices do not guarantee lower enterprise AI bills, and how organizations can measure and manage cost per business outcome as adoption scales.
Deploying Quantized Gemma 3 on AI Edge Infrastructure: Low-Memory LLM Efficiency
Explore how Quantization-Aware Training (QAT) enables Gemma 3 27B-IT Q4 models to run efficiently on AI edge infrastructure with low memory and fast inference.
Inside VINO: Unifying Text, Video, and Image Editing in One Architecture
A deep dive into VINO, a unified visual generation and editing framework combining a Vision-Language Model with a Multi-Modal Diffusion Transformer to handle seven distinct tasks from a single model checkpoint.
Optimizing AI Model Local Deployment: Navigating Multimodal Quantization Beyond AutoAWQ Deprecation
A practical guide to preparing fine-tuned multimodal models for local deployment, diagnosing quantization wrapper failures, and validating exports before serving them.
Linear Math and Cross-Model KV Cache Transfers Unlock Scalable AI Compute
An original article explaining how Nvidia's linear math technique eliminates the costly recomputation tax when agentic AI systems handoff tasks between models, using cross-model KV cache transfer to achieve up to 25x speedups while preserving up to 98% accuracy.